<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/">
    <channel>
        <title>jongkhurun’s blog</title>
        <link>https://blog.jongkhu.com//</link>
        <description>MILab打工人的论文阅读笔记</description>
        <lastBuildDate>Sun, 02 Feb 2025 15:24:11 GMT</lastBuildDate>
        <docs>https://validator.w3.org/feed/docs/rss2.html</docs>
        <generator>https://github.com/jpmonette/feed</generator>
        <language>zh-CN</language>
        <copyright>All rights reserved 2025, jongkhu29</copyright>
        <item>
            <title><![CDATA[Universal Image Restoration Pre-training via Degradation Classification]]></title>
            <link>https://blog.jongkhu.com//article/dcpt</link>
            <guid>https://blog.jongkhu.com//article/dcpt</guid>
            <pubDate>Sun, 26 Jan 2025 00:00:00 GMT</pubDate>
            <description><![CDATA[Accepted by ICLR 2025.]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-996763ecec6346f1990b94a2fbcd9df2"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-1874b79d7c2c8066a787dc01b6332924"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://raw.githubusercontent.com/jkhu29/jkhu29/refs/heads/main/WechatIMG840.jpg?t=1874b79d-7c2c-8066-a787-dc01b6332924" alt="notion image" loading="lazy" decoding="async"/></div></figure><div class="notion-text notion-block-c44b651d1a334868a87d7cdd112cb6fa">论文地址：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://openreview.net/forum?id=PacBhLzeGO">https://openreview.net/forum?id=PacBhLzeGO</a></div><div class="notion-text notion-block-87df4ed70e3c4fc390fe8dc06dd64705">代码地址：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/MILab-PKU/dcpt">https://github.com/MILab-PKU/dcpt</a></div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-2265ce5bf7844b659c59345910f69ae6" data-id="2265ce5bf7844b659c59345910f69ae6"><span><div id="2265ce5bf7844b659c59345910f69ae6" class="notion-header-anchor"></div><a class="notion-hash-link" href="#2265ce5bf7844b659c59345910f69ae6" title="背景"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">背景</span></span></h3><div class="notion-text notion-block-075f0fbc321f4d5c8bbf64df67cb9ab3">图像复原是利用模型将低质量（LQ）图像改进为高质量（HQ）图像的任务，在深度学习时代，图像复原任务可以被进一步理解为：<b>以低质量图像为条件生成高质量图像</b>。</div><div class="notion-text notion-block-2c61624b59b94501810c6390f94d884e">通用图像复原（Universal Image Restoration, UIR）任务是图像复原的一项重要的子任务。UIR 试图创造一种方法，使得模型能够自主的应对不同退化，并生成语义、细节纹理一致的高质量图像。可以简单的认为，一个合格的UIR模型应当包含以下两种能力：</div><div class="notion-text notion-block-750d2d8af57741c58835e89133a101c9">这导向了两种不同的通用图像复原方法设计思路：（1）促进退化判别；（2）引入生成Prior。其中前者已经被得到广泛的研究。流行的方法使用输入图像的退化表征作为判别提示，如：梯度、频率、附加参数和经神经网络压缩的抽象特征等等。虽然这些方法通过使用精确有效的退化提示获得了很高的复原性能，但它们<b>未能利用复原模型本身所蕴含的潜在先验信息</b>。</div><div class="notion-text notion-block-1b3a473b8471463bad7b8e8274829457">DCPT的诞生来源于对复原模型自我退化判别能力的分析。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-ad1af613633547e4a42021e30902ce8c" data-id="ad1af613633547e4a42021e30902ce8c"><span><div id="ad1af613633547e4a42021e30902ce8c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#ad1af613633547e4a42021e30902ce8c" title="发现"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">发现</span></span></h3><div class="notion-text notion-block-d092980d19124d879872045d67784dff">我们对复原模型自身的退化判别能力进行了分析，并得到三个有趣的发现：</div><ol start="1" class="notion-list notion-list-numbered notion-block-8e880bc90c0748b29ff651f2bc7e92c1"><li>随机初始化模型显示出对退化进行分类的内在能力；</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-0710ad4e27804b3c9d8bcd03f8c53865"><li>在一体化（All-in-one）复原任务中训练的模型表现出辨别未知退化的能力；</li></ol><ol start="3" class="notion-list notion-list-numbered notion-block-69b7479a18b94483a7c4aae1a4c09e93"><li>在修复模型的早期训练中，有一个退化理解步骤。</li></ol><div class="notion-text notion-block-c5f62a7ec04c4b95af1e662e945490f9">我们进行了一项简单的预实验来说明这三点：我们提取了复原训练过程中网络复原头之前的输出特征，训练过程中，模型仅见到雾霾、雨、高斯噪声三种退化。根据该特征， kNN 分类器将对五种退化类型（包括雾霾、雨天、高斯噪声、运动模糊和弱光）进行分类。</div><div class="notion-text notion-block-21171a85afb44e9191d6da1669cb5384">预实验结果如下：</div><table class="notion-simple-table notion-block-a066b0e041394a3f87ecd8a471632f91"><tbody><tr class="notion-simple-table-row notion-block-3ce7dc8501684ad4b832160fc25aa31c"><td class="" style="width:120px"><div class="notion-simple-table-cell">Methods</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">NAFNet</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">SwinIR</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Restormer</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">PromptIR</div></td></tr><tr class="notion-simple-table-row notion-block-d8f90c6011994081b53200c35f3d28a0"><td class="" style="width:120px"><div class="notion-simple-table-cell">Acc. on Random initialized (%)</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">52 ± 1</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">64 ± 4</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">71 ± 4</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">55 ± 3</div></td></tr><tr class="notion-simple-table-row notion-block-8ec0842223854d679053fa2941738a45"><td class="" style="width:120px"><div class="notion-simple-table-cell">Acc. on 3D all-in-one trained 200k iterations (%)</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">90 ± 5</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">92 ± 6 </div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">93 ± 3 </div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">93 ± 5</div></td></tr><tr class="notion-simple-table-row notion-block-a84f8cadbd3945c78525977a5d0c1ff6"><td class="" style="width:120px"><div class="notion-simple-table-cell">Acc. on 3D all-in-one trained 400k iterations (%)</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">94 ± 4</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">95 ± 4 </div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">95 ± 4 </div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">95 ± 4</div></td></tr><tr class="notion-simple-table-row notion-block-c993b820e3424b7681a32010e9147b97"><td class="" style="width:120px"><div class="notion-simple-table-cell">Acc. on 3D all-in-one trained 600k iterations (%)</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">94 ± 5</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">95 ± 4</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">97 ± 2 </div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">95 ± 4</div></td></tr></tbody></table><div class="notion-text notion-block-71e043cd64a7428e861119a2db075e91">可以看到四种网络在网络初始化时就表现出52%～71%的分类准确率，且在复原训练过程早期（前200k次迭代）快速收敛到90%以上的分类准确率。</div><details class="notion-toggle notion-block-e001eb59fce64b9fa5ee198aea32422e"><summary>当退化数量进一步增多，…</summary><div><div class="notion-text notion-block-84cbc18832db4a8f8470cbefe1230870">遗憾的是，我们发现复原模型对未知退化的辨别能力会随着退化种类的增多而逐渐减弱。我们将在后续工作中对此进行更充分的讨论。</div></div></details><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-1884b79d7c2c8040a8aff97de510dbce" data-id="1884b79d7c2c8040a8aff97de510dbce"><span><div id="1884b79d7c2c8040a8aff97de510dbce" class="notion-header-anchor"></div><a class="notion-hash-link" href="#1884b79d7c2c8040a8aff97de510dbce" title="动机"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">动机</span></span></h4><div class="notion-text notion-block-5940da46bf30459f91318ecedc663a85">由于图像复原的核心任务还是以低质量图像为条件生成高质量图像，我们不希望在复原训练过程中出现与该任务存在潜在冲突的其他训练子任务，例如退化分类。于是，我们选择将显式地将该训练阶段提前为“预训练”，并进一步创造了DCPT。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-3da79ada3df740beae12c5ebc67670f7" data-id="3da79ada3df740beae12c5ebc67670f7"><span><div id="3da79ada3df740beae12c5ebc67670f7" class="notion-header-anchor"></div><a class="notion-hash-link" href="#3da79ada3df740beae12c5ebc67670f7" title="方法"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">方法</span></span></h3><div class="notion-text notion-block-b042477596cf4739918518306c704617"><b>Degradation Classification Pre-Training </b>(DCPT) 是一个简单且有效的方法，可见下图。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-1874b79d7c2c80c0ac2ee80f8cc769c9"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://raw.githubusercontent.com/jkhu29/jkhu29/refs/heads/main/WechatIMG841.jpg?t=1874b79d-7c2c-80c0-ac2e-e80f8cc769c9" alt="notion image" loading="lazy" decoding="async"/></div></figure><div class="notion-text notion-block-7f62875fc66d406e9b4dc025278fdf04">在单次迭代中，它包含两个阶段：退化分类阶段、生成阶段，这两个阶段交替进行。其中，</div><ul class="notion-list notion-list-disc notion-block-63e65144cf4c44e288fa8592a91138e5"><li>退化分类阶段：通过提取复原网络的深层特征，并将其输入一个轻量级分类器，以对输入图像的退化种类进行分类。</li></ul><ul class="notion-list notion-list-disc notion-block-3a6b9bb3f35349f18f844e545caea2b6"><li>生成阶段：我们利用最原始的Autoencoder手段对复原模型的生成能力进行保留。</li></ul><div class="notion-text notion-block-55678141875141e48aea97489e0639c1">实现代码也非常简洁：</div><div class="notion-text notion-block-0809595d72d849f6bc858b3d25371a24">需要注意，在预训练结束后，仍需要进行复原任务上的fine-tune。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-917ad9c6456044868b286bd581572ff4" data-id="917ad9c6456044868b286bd581572ff4"><span><div id="917ad9c6456044868b286bd581572ff4" class="notion-header-anchor"></div><a class="notion-hash-link" href="#917ad9c6456044868b286bd581572ff4" title="实验结果"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">实验结果</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-fd2c18cd6e2b4f379bf26eee7ebcaaa0" data-id="fd2c18cd6e2b4f379bf26eee7ebcaaa0"><span><div id="fd2c18cd6e2b4f379bf26eee7ebcaaa0" class="notion-header-anchor"></div><a class="notion-hash-link" href="#fd2c18cd6e2b4f379bf26eee7ebcaaa0" title="5D All-in-one image restoration"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">5D All-in-one image restoration</span></span></h4><table class="notion-simple-table notion-block-456504d9c0be4356bb244a018ed07cac"><tbody><tr class="notion-simple-table-row notion-block-1e09b30318a84da0a0d298405f8220c8"><td class="" style="width:169px"><div class="notion-simple-table-cell">Method</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Dehazing</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Deraining</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Denoising</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Deblurring</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Low-Light</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Average</div></td></tr><tr class="notion-simple-table-row notion-block-d6c656529c10479aaffbdc14aacd0348"><td class="" style="width:169px"><div class="notion-simple-table-cell">ㅤ</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">on SOTS</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">on Test100L</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">on BSD68</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">on GoPro</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">on LOL</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">ㅤ</div></td></tr><tr class="notion-simple-table-row notion-block-e3ada440aaf24cf5a682c8e065e54973"><td class="" style="width:169px"><div class="notion-simple-table-cell">ㅤ</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">PSNR / SSIM</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">PSNR / SSIM</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">PSNR / SSIM</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">PSNR / SSIM</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">PSNR / SSIM</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">PSNR / SSIM</div></td></tr><tr class="notion-simple-table-row notion-block-7d50b6a8787149e8867b59d3d0fecacc"><td class="" style="width:169px"><div class="notion-simple-table-cell">AirNet</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">21.04 / 0.884</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">32.98 / 0.951</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">30.91 / 0.882</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">24.35 / 0.781</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">18.18 / 0.735</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">25.49 / 0.846</div></td></tr><tr class="notion-simple-table-row notion-block-771ef66734bd4421a4f5d3f02de04faa"><td class="" style="width:169px"><div class="notion-simple-table-cell">IDR</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">25.24 / 0.943</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">35.63 / 0.965</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">31.60 / 0.887</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">27.87 / 0.846</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">21.34 / 0.826</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">28.34 / 0.893</div></td></tr><tr class="notion-simple-table-row notion-block-da38503d3c364540bcfab0cc1736a95f"><td class="" style="width:169px"><div class="notion-simple-table-cell">InstructIR</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">27.10 / 0.956</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">36.84 / 0.973</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">31.40 / 0.887</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">29.40 / 0.886</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">23.00 / 0.836</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">29.55 / 0.907</div></td></tr><tr class="notion-simple-table-row notion-block-6f55a94aad47444dbf513ef4f426addb"><td class="" style="width:169px"><div class="notion-simple-table-cell">SwinIR</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">21.50 / 0.891</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">30.78 / 0.923</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">30.59 / 0.868</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">24.52 / 0.773</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">17.81 / 0.723</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">25.04 / 0.835</div></td></tr><tr class="notion-simple-table-row notion-block-04d29cb241924346ae8e5d03f84f3366"><td class="" style="width:169px"><div class="notion-simple-table-cell"><b>DCPT-SwinIR</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>28.67</b> / <b>0.973</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>35.70</b> / <b>0.974</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>31.16</b> / <b>0.882</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>26.42</b> / <b>0.807</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>20.38</b> / <b>0.836</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>28.47</b> / <b>0.894</b></div></td></tr><tr class="notion-simple-table-row notion-block-bd2729c592054db38ea8d6e3a4015446"><td class="" style="width:169px"><div class="notion-simple-table-cell">NAFNet</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">25.23 / 0.939</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">35.56 / 0.967</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">31.02 / 0.883</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">26.53 / 0.808</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">20.49 / 0.809</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">27.76 / 0.881</div></td></tr><tr class="notion-simple-table-row notion-block-7b3bdf92245b4abdaa217d1fde1c4472"><td class="" style="width:169px"><div class="notion-simple-table-cell"><b>DCPT-NAFNet</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>29.47</b> / <b>0.971</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>35.68</b> / <b>0.973</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>31.31</b> / <b>0.886</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>29.22</b> / <b>0.883</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>23.52</b> / <b>0.855</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>29.84</b> / <b>0.914</b></div></td></tr><tr class="notion-simple-table-row notion-block-398c1ead30f842c2b98f3f00cf88aa2b"><td class="" style="width:169px"><div class="notion-simple-table-cell">Restormer</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">24.09 / 0.927</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">34.81 / 0.962</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">31.49 / 0.884</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">27.22 / 0.829</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">20.41 / 0.806</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">27.60 / 0.881</div></td></tr><tr class="notion-simple-table-row notion-block-ec7bc4d0dd71483c97b2ab25271754a7"><td class="" style="width:169px"><div class="notion-simple-table-cell"><b>DCPT-Restormer</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>29.86</b> / <b>0.973</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>36.68</b> / <b>0.975</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>31.46</b> / <b>0.888</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>28.95</b> / <b>0.879</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>23.26</b> / <b>0.842</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>30.04</b> / <b>0.911</b></div></td></tr><tr class="notion-simple-table-row notion-block-f47a1b3d97094a029c0c2986f8b1d610"><td class="" style="width:169px"><div class="notion-simple-table-cell">PromptIR</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">25.20 / 0.931</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">35.94 / 0.964</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">31.17 / 0.882</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">27.32 / 0.842</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">20.94 / 0.799</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">28.11 / 0.883</div></td></tr><tr class="notion-simple-table-row notion-block-db567f6f1e454326875de73fff592098"><td class="" style="width:169px"><div class="notion-simple-table-cell"><b>DCPT-PromptIR</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>30.72</b> / <b>0.977</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>37.32</b> / <b>0.978</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>31.32</b> / <b>0.885</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>28.84</b> / <b>0.877</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>23.35</b> / <b>0.840</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>30.31</b> / <b>0.911</b></div></td></tr></tbody></table><div class="notion-text notion-block-7b60a51330924a3da749cbbf1bbd8439">可以看出，无论是 CNN 网络还是 Transformer 网络，无论是直线网络还是类 UNet 网络，DCPT在 5D All-in-one image restoration 任务上的平均性能提升始终保持在 2.08 dB 及以上。</div><div class="notion-text notion-block-d292fc4a955d405f907fc965adb37fed">我们也展示一些可视化数据，以证明DCPT也确实能提升输出图像的视觉感官。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-1874b79d7c2c808985f1f5fe6f7f59c6"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://raw.githubusercontent.com/jkhu29/jkhu29/refs/heads/main/WechatIMG842.jpg?t=1874b79d-7c2c-8089-85f1-f5fe6f7f59c6" alt="notion image" loading="lazy" decoding="async"/></div></figure><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-2080ae14c9104d29bf3130ade62fc153" data-id="2080ae14c9104d29bf3130ade62fc153"><span><div id="2080ae14c9104d29bf3130ade62fc153" class="notion-header-anchor"></div><a class="notion-hash-link" href="#2080ae14c9104d29bf3130ade62fc153" title="10D All-in-one image restoration"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">10D All-in-one image restoration</span></span></h4><div class="notion-text notion-block-1679900a87a8429d9ca430e9e7b0b458">我们选取了十种退化进行试验，并绘制了雷达图。</div><div class="notion-row notion-block-1874b79d7c2c8092899df3e6a78117f1"><div class="notion-column notion-block-1874b79d7c2c802aa225eee80f24190d" style="width:calc((100% - (1 * min(32px, 4vw))) * 0.5)"><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-1874b79d7c2c80ad8508c83df9216006"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:336px;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://raw.githubusercontent.com/jkhu29/jkhu29/refs/heads/main/dcpt_10d_PSNR.png?t=1874b79d-7c2c-80ad-8508-c83df9216006" alt="PSNR" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">PSNR</figcaption></div></figure></div><div class="notion-spacer"></div><div class="notion-column notion-block-1874b79d7c2c8032bad4e73ea36e229c" style="width:calc((100% - (1 * min(32px, 4vw))) * 0.5)"><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-1874b79d7c2c80bbbf34c5f99ae68911"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://raw.githubusercontent.com/jkhu29/jkhu29/refs/heads/main/dcpt_10d_SSIM.png?t=1874b79d-7c2c-80bb-bf34-c5f99ae68911" alt="SSIM" loading="lazy" decoding="async"/><figcaption class="notion-asset-caption">SSIM</figcaption></div></figure></div><div class="notion-spacer"></div></div><div class="notion-text notion-block-bf6ea17675994a7da962b5067f733af0">可以看到使用DCPT预训练后，NAFNet的PSNR与SSIM指标都有比较显著的提升。具体数值指标如下：</div><table class="notion-simple-table notion-block-63987a5d37e4471287707d58a1748da9"><tbody><tr class="notion-simple-table-row notion-block-71197ff8fcaa47738faf3d99b16e87e9"><td class="" style="width:120px"><div class="notion-simple-table-cell">Method</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">Average</div></td></tr><tr class="notion-simple-table-row notion-block-5ad684684ff74a789e5051023f86cfd3"><td class="" style="width:120px"><div class="notion-simple-table-cell">AirNet</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">26.41 / 0.842</div></td></tr><tr class="notion-simple-table-row notion-block-be70cdc6964d463fb1199b56c58cc47b"><td class="" style="width:120px"><div class="notion-simple-table-cell">TransWeather</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">22.83 / 0.779</div></td></tr><tr class="notion-simple-table-row notion-block-8d40324f3d1f4c56b7eac7308c9ee721"><td class="" style="width:120px"><div class="notion-simple-table-cell">WeatherDiff</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">24.60 / 0.793</div></td></tr><tr class="notion-simple-table-row notion-block-7c6fcd754cea4ac69c85f5d224ae8151"><td class="" style="width:120px"><div class="notion-simple-table-cell">PromptIR</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">27.93 / 0.851</div></td></tr><tr class="notion-simple-table-row notion-block-e6cdf5f44e9740f8826125131f77ff87"><td class="" style="width:120px"><div class="notion-simple-table-cell">DiffUIR-L</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">28.75 / 0.869</div></td></tr><tr class="notion-simple-table-row notion-block-fe7ece607d884222af249bec48d776b4"><td class="" style="width:120px"><div class="notion-simple-table-cell">NAFNet</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">27.17 / 0.837</div></td></tr><tr class="notion-simple-table-row notion-block-9c88976b81934f03b986bb766c22adc9"><td class="" style="width:120px"><div class="notion-simple-table-cell">    + DACLIP</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">27.42 / 0.798</div></td></tr><tr class="notion-simple-table-row notion-block-a11bb7722f914410ae9f9f2d84af7506"><td class="" style="width:120px"><div class="notion-simple-table-cell">    + Instruct</div></td><td class="" style="width:120px"><div class="notion-simple-table-cell">28.30 / 0.862</div></td></tr><tr class="notion-simple-table-row notion-block-df9b8b77b30847b9964e9f74393fc300"><td class="" style="width:120px"><div class="notion-simple-table-cell">    + <b>DCPT (Ours)</b></div></td><td class="" style="width:120px"><div class="notion-simple-table-cell"><b>29.72 / 0.888</b></div></td></tr></tbody></table><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-3a0bf6727ce1450a97941f5557f6816f" data-id="3a0bf6727ce1450a97941f5557f6816f"><span><div id="3a0bf6727ce1450a97941f5557f6816f" class="notion-header-anchor"></div><a class="notion-hash-link" href="#3a0bf6727ce1450a97941f5557f6816f" title="Transfer learning"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Transfer learning</span></span></h4><div class="notion-text notion-block-710b0626de934c2ca2a7b8ec10e1560e">众所周知，图像复原模型的过拟合现象严重。在A退化任务下训练的复原模型极难泛化到B退化任务。我们发现DCPT种的退化分类器有助于模型跨任务泛化。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-1874b79d7c2c806db096cd83aa8113f1"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://raw.githubusercontent.com/jkhu29/jkhu29/refs/heads/main/WechatIMG843.jpg?t=1874b79d-7c2c-806d-b096-cd83aa8113f1" alt="notion image" loading="lazy" decoding="async"/></div></figure><div class="notion-text notion-block-33c02170f61a4c309d3dcb4c5512c4e9">更多其他实验结果请关注我们的<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://openreview.net/forum?id=PacBhLzeGO">文章</a>。</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-6961a284309b442cad7d3fe6b71ea78c" data-id="6961a284309b442cad7d3fe6b71ea78c"><span><div id="6961a284309b442cad7d3fe6b71ea78c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#6961a284309b442cad7d3fe6b71ea78c" title="讨论"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">讨论</span></span></h3><ul class="notion-list notion-list-disc notion-block-726ff585fe4a4bc9b2f7527a60c8b7a2"><li>Q: 在复原训练中，仅仅只是存在退化分类嘛？</li></ul><div class="notion-text notion-block-39f5734bd9874a269db5ac02f0494ce1">A. 否定。或许可以在实验结果中窥见一二。在5D All-in-one任务中，SwinIR的去雾结果展示出明显的“区域性”。我们猜测，对于退化在全局图像中不均匀出现的情况，复原模型也会对输入图像的退化进行“分割”。</div><ul class="notion-list notion-list-disc notion-block-1f84302c71814ccbb13ade98fedb660d"><li><b>复原中隐藏着（退化）辨别</b></li></ul><div class="notion-text notion-block-cc38be3e28da4e74b95e72c6109d6114">之前的<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://arxiv.org/abs/2108.00406">研究</a>已经调查了超分辨率模型在复原过程中区分不同类型退化的能力。DCPT的初步实验也表明，随机初始化模型能够对退化进行分类。此外，一体化复原训练也能增强了模型的退化分类能力，并赋予复原模型在退化分类任务上的泛化能力。这些结果表明，<b>复原中隐藏着（退化）辨别</b>。</div><div class="notion-text notion-block-144a52329c2046d092ac5f8a4abbcd40">DCPT的实验结果凸显了判别先验在图像复原预训练中的有效性。这些结果表明，在训练前将足够的判别信息纳入模型可以显著提高其性能。我们假设，在复原模型中加入卓越的降解感知判别信息，并最大限度地提高其判别能力，将进一步提高模型的复原性能。预计这一假设将为通用复原领域开发大量新型预训练方法铺平道路。</div></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Debug about fused shift-window]]></title>
            <link>https://blog.jongkhu.com//article/debug-shift-window</link>
            <guid>https://blog.jongkhu.com//article/debug-shift-window</guid>
            <pubDate>Sun, 03 Dec 2023 00:00:00 GMT</pubDate>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-1e57842d12834e5ba9e3ea853db4e6bc"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><div class="notion-text notion-block-91c85bac93ae4565a28bae1a7900c30d">在试图白嫖<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/microsoft/Swin-Transformer/tree/main/kernels/window_process">FuseWindowProcess</a>时，发现window_partition的前向没有问题，但是window_merge的前向在推理过程中经常容易爆错，本篇博客用于解决一下这个问题。</div><div class="notion-text notion-block-f69ec960557e4564ba33d05c34e1d6d1">出事的代码如下：</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-bf0b1af73b474b7ab2cc2bac565b503f" data-id="bf0b1af73b474b7ab2cc2bac565b503f"><span><div id="bf0b1af73b474b7ab2cc2bac565b503f" class="notion-header-anchor"></div><a class="notion-hash-link" href="#bf0b1af73b474b7ab2cc2bac565b503f" title="问题描述与定位"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">问题描述与定位</span></span></h3><div class="notion-text notion-block-1241c82d865249a8a05be68cfd094119">上面这串代码会在一些输入分辨率非方形的情况下（例如分辨率在256*128）在推理过程中爆错。</div><h4 class="notion-h notion-h3 notion-h-indent-1 notion-block-24ac443eb16941ca8ebd8341b5a3f6c1" data-id="24ac443eb16941ca8ebd8341b5a3f6c1"><span><div id="24ac443eb16941ca8ebd8341b5a3f6c1" class="notion-header-anchor"></div><a class="notion-hash-link" href="#24ac443eb16941ca8ebd8341b5a3f6c1" title="Debug一下"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Debug一下</span></span></h4><div class="notion-text notion-block-4aad5d9cb75e490b961d2d067780853f">我们首先编译一下debug版本cuda然后进入cuda-gdb</div><div class="notion-text notion-block-5bf94fcfb7ee47c4bedf436adcf4fcc9">GDB，启动！</div><div class="notion-text notion-block-06f98ac9d4414763bfc7a937c5de9cbc">按r运行之后出现这样一个问题。</div><div class="notion-text notion-block-cb2340c81f6b4e01b9da9f3c1d11192d">bt一下：</div><div class="notion-text notion-block-dc94f0c0d226410a9b288a8a34808815">这里有一个函数：__ldg，Load Global Device，他的作用是：将全局内存中的数据放入缓存中，从而提高读取速度并确保数据的一致性。在<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#global-memory-3-0">cuda文档</a>中，有其详细的描述：</div><blockquote class="notion-quote notion-block-4af5ba6751af40ff9e3a0706d220a5bb"><div>Data that is read-only for the entire lifetime of the kernel can also be cached in the read-only data cache described in the previous section by reading it using the __ldg() function (see Read-Only Data Cache Load Function). When the compiler detects that the read-only condition is satisfied for some data, it will use __ldg() to read it. The compiler might not always be able to detect that the read-only condition is satisfied for some data. Marking pointers used for loading such data with both the const and <b>restrict</b> qualifiers increases the likelihood that the compiler will detect the read-only condition.</div></blockquote><div class="notion-text notion-block-f0bf66949ad841f1b83f09ab9a668017">一般来说__ldg不会有问题，input是输入，output是输出，我们先排除这俩的问题，比如尽量避免出现nan。我们在WindowProcessReverse调用前后都添加上assert：</div><div class="notion-text notion-block-7857c7dc48044e4c8b715ff16aa3c4bf">运行之后报错，但是报错语句出现在reshape后：</div><div class="notion-text notion-block-fd751123af9445f393fdbd7115ad4fe5">这就很诡异了，这意味着WindowProcessReverse的输出是正常的，但是经过了reshape后出现了OOM问题。目前不清楚是reshape的问题还是x自己的问题，于是我多尝试了几种：</div><div class="notion-text notion-block-235a4c5e86274c04902c7c5d744a728b">报错如下：</div><div class="notion-text notion-block-16d58987d75e4da0ba079eac865788cd">可以见到，对x进行任何操作，包括inplace操作，都会导致illegal memory access，也就是说x本身可能指向了一些非法内存，定位到下面这行语句：</div><h3 class="notion-h notion-h2 notion-h-indent-0 notion-block-9e0c4c3f6c1747d6ac41c5036d270f40" data-id="9e0c4c3f6c1747d6ac41c5036d270f40"><span><div id="9e0c4c3f6c1747d6ac41c5036d270f40" class="notion-header-anchor"></div><a class="notion-hash-link" href="#9e0c4c3f6c1747d6ac41c5036d270f40" title="解决问题"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">解决问题</span></span></h3><div class="notion-text notion-block-eb605d992f7142ffa8c679caebb20c90">原来就是这个地方应该对宽nW进行处理，原来大佬有时候也会犯这种小错误。</div><div class="notion-blank notion-block-7dcaad47a1564499a38b97e602c58ddc"> </div><div class="notion-text notion-block-b28774fd636341878fe3e6facde714ec">散会！</div></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Event driven image and video processing dataset]]></title>
            <link>https://blog.jongkhu.com//article/event-processing</link>
            <guid>https://blog.jongkhu.com//article/event-processing</guid>
            <pubDate>Mon, 24 Jun 2024 00:00:00 GMT</pubDate>
            <description><![CDATA[使用event做辅助的图像或视频处理数据集整理]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-4962bed0f1fc44bea9b65caa085c6999"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><div class="notion-text notion-block-8349c2f581134706b4f009774a84258b">将主要整理使用event作为输入的图像、视频的重建、复原、增强相关数据集。</div><h2 class="notion-h notion-h1 notion-h-indent-0 notion-block-3ffb065581cf42519a819815b3b9dd92" data-id="3ffb065581cf42519a819815b3b9dd92"><span><div id="3ffb065581cf42519a819815b3b9dd92" class="notion-header-anchor"></div><a class="notion-hash-link" href="#3ffb065581cf42519a819815b3b9dd92" title="概念明晰"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">概念明晰</span></span></h2><div class="notion-text notion-block-0aac574a33d74148b09ec159d2f5a7ed">首先给出本blog中使用的一些概念术语。</div><div class="notion-text notion-block-ac0363e63de04500b0d275d015c47bfc">任务的定义：</div><ul class="notion-list notion-list-disc notion-block-c04727dabbb542dd9915cd2317ca2a37"><li>Reconstruction: 一般的图像、视频重建以及HDR重建。</li></ul><ul class="notion-list notion-list-disc notion-block-bac40545a6b84e749d4ff5170e3553af"><li>Restoration: 指将恶劣环境（暗光、抖动、噪声等）场景下拍摄得到的图像或视频转化为干净、锐利的图像或视频。</li></ul><ul class="notion-list notion-list-disc notion-block-3b92bf1964964574b51b195edf288b5a"><li>Enhancement: </li></ul><ul class="notion-list notion-list-disc notion-block-09a3b46272c24be587e2b0f821543c40"><li>Super Resolution: 指将图像、视频的空间分辨率提高。注意，在本篇blog，Super-Resolution不包含在Restoration任务中。</li></ul><ul class="notion-list notion-list-disc notion-block-d1812626e98448359ef3d2c2adc1415b"><li>Frame Interpolation: 指插帧。例如，将视频的帧率从15fps提升到60fps。</li></ul><div class="notion-text notion-block-50ecf6d17e964ee8b6f9fe1f76f0a5ea">任务的性质：</div><ul class="notion-list notion-list-disc notion-block-b10fd2551d5c4edcb941ef5a983be78c"><li>Event-based: 仅使用事件（Event）数据作为输入的方法；</li></ul><ul class="notion-list notion-list-disc notion-block-6de317f656394de28955d39c29e88a50"><li>Event-guided: 使用事件（Event）与帧（Frame）共同输入的方法；</li></ul><ul class="notion-list notion-list-disc notion-block-b70c2df755394884a0b32202bdb3bcc2"><li>Event-driven: 上述两种方法的并集。</li></ul><blockquote class="notion-quote notion-block-890676642a3649918d2ffb245e277c89"><div>一般来说，Event-guided为主流。</div></blockquote><h2 class="notion-h notion-h1 notion-h-indent-0 notion-block-79951b0a70804ebf8374b0c89d897898" data-id="79951b0a70804ebf8374b0c89d897898"><span><div id="79951b0a70804ebf8374b0c89d897898" class="notion-header-anchor"></div><a class="notion-hash-link" href="#79951b0a70804ebf8374b0c89d897898" title="Image"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Image</span></span></h2><div class="notion-text notion-block-0c65f7d30bd44879bbf6d02ec31a67b7">本节中将主要介绍使用事件数据进行图像处理的数据集。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-a0b57018a699429bb27eb19faaecf70d" data-id="a0b57018a699429bb27eb19faaecf70d"><span><div id="a0b57018a699429bb27eb19faaecf70d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a0b57018a699429bb27eb19faaecf70d" title="Image reconstruction"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Image reconstruction</span></span></h3><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-f84df5fe9c764b7f9a4f8a0b02795a2e" data-id="f84df5fe9c764b7f9a4f8a0b02795a2e"><span><div id="f84df5fe9c764b7f9a4f8a0b02795a2e" class="notion-header-anchor"></div><a class="notion-hash-link" href="#f84df5fe9c764b7f9a4f8a0b02795a2e" title="Image Restoration"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Image Restoration</span></span></h3><h2 class="notion-h notion-h1 notion-h-indent-0 notion-block-2ed3271faca94a0b8f2296b6122314a2" data-id="2ed3271faca94a0b8f2296b6122314a2"><span><div id="2ed3271faca94a0b8f2296b6122314a2" class="notion-header-anchor"></div><a class="notion-hash-link" href="#2ed3271faca94a0b8f2296b6122314a2" title="Video"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Video</span></span></h2><div class="notion-text notion-block-4403c3ad059c4ac8aa747c4f73743e25">本节中将主要介绍使用事件数据进行视频处理的数据集。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-5a221b0b669b43de8d1fe192bfb7517a" data-id="5a221b0b669b43de8d1fe192bfb7517a"><span><div id="5a221b0b669b43de8d1fe192bfb7517a" class="notion-header-anchor"></div><a class="notion-hash-link" href="#5a221b0b669b43de8d1fe192bfb7517a" title="Event-to-Video reconstruction"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Event-to-Video reconstruction</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-1b127851961b42bf95b000cb10f151da" data-id="1b127851961b42bf95b000cb10f151da"><span><div id="1b127851961b42bf95b000cb10f151da" class="notion-header-anchor"></div><a class="notion-hash-link" href="#1b127851961b42bf95b000cb10f151da" title="测试数据集"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title"><b>测试数据集</b></span></span></h4><div class="notion-text notion-block-c79f1ebed75b4aa1b56ca7c3eca61bcf">首推一个非常完善的测评库：<a target="_blank" rel="noopener noreferrer" href="https://github.com/ercanburak/EVREAL" class="notion-external notion-external-mention"><div class="notion-external-image"><svg viewBox="0 0 260 260"><g><path d="M128.00106,0 C57.3172926,0 0,57.3066942 0,128.00106 C0,184.555281 36.6761997,232.535542 87.534937,249.460899 C93.9320223,250.645779 96.280588,246.684165 96.280588,243.303333 C96.280588,240.251045 96.1618878,230.167899 96.106777,219.472176 C60.4967585,227.215235 52.9826207,204.369712 52.9826207,204.369712 C47.1599584,189.574598 38.770408,185.640538 38.770408,185.640538 C27.1568785,177.696113 39.6458206,177.859325 39.6458206,177.859325 C52.4993419,178.762293 59.267365,191.04987 59.267365,191.04987 C70.6837675,210.618423 89.2115753,204.961093 96.5158685,201.690482 C97.6647155,193.417512 100.981959,187.77078 104.642583,184.574357 C76.211799,181.33766 46.324819,170.362144 46.324819,121.315702 C46.324819,107.340889 51.3250588,95.9223682 59.5132437,86.9583937 C58.1842268,83.7344152 53.8029229,70.715562 60.7532354,53.0843636 C60.7532354,53.0843636 71.5019501,49.6441813 95.9626412,66.2049595 C106.172967,63.368876 117.123047,61.9465949 128.00106,61.8978432 C138.879073,61.9465949 149.837632,63.368876 160.067033,66.2049595 C184.49805,49.6441813 195.231926,53.0843636 195.231926,53.0843636 C202.199197,70.715562 197.815773,83.7344152 196.486756,86.9583937 C204.694018,95.9223682 209.660343,107.340889 209.660343,121.315702 C209.660343,170.478725 179.716133,181.303747 151.213281,184.472614 C155.80443,188.444828 159.895342,196.234518 159.895342,208.176593 C159.895342,225.303317 159.746968,239.087361 159.746968,243.303333 C159.746968,246.709601 162.05102,250.70089 168.53925,249.443941 C219.370432,232.499507 256,184.536204 256,128.00106 C256,57.3066942 198.691187,0 128.00106,0 Z M47.9405593,182.340212 C47.6586465,182.976105 46.6581745,183.166873 45.7467277,182.730227 C44.8183235,182.312656 44.2968914,181.445722 44.5978808,180.80771 C44.8734344,180.152739 45.876026,179.97045 46.8023103,180.409216 C47.7328342,180.826786 48.2627451,181.702199 47.9405593,182.340212 Z M54.2367892,187.958254 C53.6263318,188.524199 52.4329723,188.261363 51.6232682,187.366874 C50.7860088,186.474504 50.6291553,185.281144 51.2480912,184.70672 C51.8776254,184.140775 53.0349512,184.405731 53.8743302,185.298101 C54.7115892,186.201069 54.8748019,187.38595 54.2367892,187.958254 Z M58.5562413,195.146347 C57.7719732,195.691096 56.4895886,195.180261 55.6968417,194.042013 C54.9125733,192.903764 54.9125733,191.538713 55.713799,190.991845 C56.5086651,190.444977 57.7719732,190.936735 58.5753181,192.066505 C59.3574669,193.22383 59.3574669,194.58888 58.5562413,195.146347 Z M65.8613592,203.471174 C65.1597571,204.244846 63.6654083,204.03712 62.5716717,202.981538 C61.4524999,201.94927 61.1409122,200.484596 61.8446341,199.710926 C62.5547146,198.935137 64.0575422,199.15346 65.1597571,200.200564 C66.2704506,201.230712 66.6095936,202.705984 65.8613592,203.471174 Z M75.3025151,206.281542 C74.9930474,207.284134 73.553809,207.739857 72.1039724,207.313809 C70.6562556,206.875043 69.7087748,205.700761 70.0012857,204.687571 C70.302275,203.678621 71.7478721,203.20382 73.2083069,203.659543 C74.6539041,204.09619 75.6035048,205.261994 75.3025151,206.281542 Z M86.046947,207.473627 C86.0829806,208.529209 84.8535871,209.404622 83.3316829,209.4237 C81.8013,209.457614 80.563428,208.603398 80.5464708,207.564772 C80.5464708,206.498591 81.7483088,205.631657 83.2786917,205.606221 C84.8005962,205.576546 86.046947,206.424403 86.046947,207.473627 Z M96.6021471,207.069023 C96.7844366,208.099171 95.7267341,209.156872 94.215428,209.438785 C92.7295577,209.710099 91.3539086,209.074206 91.1652603,208.052538 C90.9808515,206.996955 92.0576306,205.939253 93.5413813,205.66582 C95.054807,205.402984 96.4092596,206.021919 96.6021471,207.069023 Z" fill="#161614"></path></g></svg></div><div class="notion-external-description"><div class="notion-external-title">EVREAL</div><div class="notion-external-subtitle"><span>ercanburak</span><span> • </span><span>Updated <!-- -->Jan 18, 2025</span></div></div></a>。它囊括了几乎所有Event-based的视频重建测试数据集，以及主流的测评指标。</div><ul class="notion-list notion-list-disc notion-block-e238cfab5e8a4d9cb701688fd6c45b19"><li>ECD，下载链接：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/ercanburak/EVREAL/blob/main/tools/download_ECD.sh">https://github.com/ercanburak/EVREAL/blob/main/tools/download_ECD.sh</a></li></ul><ul class="notion-list notion-list-disc notion-block-19479c2484774d77bdea1aecf976342f"><li>MVSEC，下载链接：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/ercanburak/EVREAL/blob/main/tools/download_MVSEC.sh">https://github.com/ercanburak/EVREAL/blob/main/tools/download_MVSEC.sh</a></li></ul><ul class="notion-list notion-list-disc notion-block-0563e4644da5462d8133de490293956a"><li>HQF，下载链接：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/ercanburak/EVREAL/blob/main/tools/download_HQF.sh">https://github.com/ercanburak/EVREAL/blob/main/tools/download_HQF.sh</a></li></ul><ul class="notion-list notion-list-disc notion-block-159cdbc0618c455d8f269912704f92e7"><li>BS-ERGB，下载链接：<!-- -->‣</li></ul><blockquote class="notion-quote notion-block-2767b4fee76a43af838991947b8cb706"><div>另有IJRR数据集，但似乎已经失效了。最新的视频重建论文均未在该数据集上进行测试。</div></blockquote><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-3c84edc54c24444f9726310eaa846ffc" data-id="3c84edc54c24444f9726310eaa846ffc"><span><div id="3c84edc54c24444f9726310eaa846ffc" class="notion-header-anchor"></div><a class="notion-hash-link" href="#3c84edc54c24444f9726310eaa846ffc" title="训练数据集"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title"><b>训练数据集</b></span></span></h4><div class="notion-text notion-block-cdeeff2c88d042018139a04f7599cbd9">目前使用的训练数据集只要来源于<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123720528.pdf">E2VID+</a>，详细的制作方法可见：<!-- -->‣<!-- -->。</div><blockquote class="notion-quote notion-block-3e62589ca5864c85a1ab56c64377c583"><div>由于需要安装ROS，制作过程可能需要一些耐心（</div></blockquote><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-be44d7965e054bed9ba2284f008248bc" data-id="be44d7965e054bed9ba2284f008248bc"><span><div id="be44d7965e054bed9ba2284f008248bc" class="notion-header-anchor"></div><a class="notion-hash-link" href="#be44d7965e054bed9ba2284f008248bc" title="推荐文献"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title"><b>推荐文献</b></span></span></h4><div class="notion-text notion-block-13b9d854cae5415187d1e1c73b9a1dd5">[1] High Speed and High Dynamic Range Video with an Event Camera, TPAMI 2020</div><div class="notion-text notion-block-3ffa8fabd67947e28e73ec78be9611ad">[2] Reducing the Sim-to-Real Gap for Event Cameras, ECCV 2020</div><div class="notion-text notion-block-dc3feaf0a5724c00a5326168557af12a">[3] Event-based Video Reconstruction Using Transformer, ICCV 2021</div><div class="notion-text notion-block-75e7cb08e2c64a3aa40fb478e46ae000">[4] Back to Event Basics: Self-Supervised Learning of Image Reconstruction for Event Cameras via Photometric Constancy, CVPR 2021</div><div class="notion-text notion-block-74d1556768b148b2bf4a00cc00e75927">[5] SPADE-E2VID: Spatially-Adaptive Denormalization for Event-Based Video Reconstruction, TIP 2021</div><div class="notion-text notion-block-6a336852edc146bbb6fc75fc0f644b4d">[6] Event-based Video Reconstruction via Potential-assisted Spiking Neural Network, CVPR 2022</div><div class="notion-text notion-block-e89a6c178a4b400e8072ca03e8f59134">[7] HyperE2VID: Improving Event-Based Video Reconstruction via Hypernetworks, TIP 2024</div><div class="notion-text notion-block-aa10ffdc1c964fea89ab9a960d9e55db">[8] Enhanced Event-Based Video Reconstruction with Motion Compensation, ECCV 2024</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-2055beea9adb4adc88caac39b913e2d6" data-id="2055beea9adb4adc88caac39b913e2d6"><span><div id="2055beea9adb4adc88caac39b913e2d6" class="notion-header-anchor"></div><a class="notion-hash-link" href="#2055beea9adb4adc88caac39b913e2d6" title="Video HDR reconstruction"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Video HDR reconstruction</span></span></h3><div class="notion-text notion-block-9e99209992b94dbb8e217b7eb7ca513b">此类任务与上述Event-to-video重建任务的区别在于，已经存在从Event-based到Event-guided的转变。</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-b2d9924d408e4bcba87dadcef6a31d87" data-id="b2d9924d408e4bcba87dadcef6a31d87"><span><div id="b2d9924d408e4bcba87dadcef6a31d87" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b2d9924d408e4bcba87dadcef6a31d87" title="数据集"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">数据集</span></span></h4><ul class="notion-list notion-list-disc notion-block-c43a28df3ca5461689a8796ba0b7b4e3"><li>Event based: HDR from ETH，下载链接：<a target="_blank" rel="noopener noreferrer" href="https://github.com/ercanburak/EVREAL/blob/main/tools/download_TPAMI20_HDR.sh" class="notion-external notion-external-mention"><div class="notion-external-image"><svg viewBox="0 0 260 260"><g><path d="M128.00106,0 C57.3172926,0 0,57.3066942 0,128.00106 C0,184.555281 36.6761997,232.535542 87.534937,249.460899 C93.9320223,250.645779 96.280588,246.684165 96.280588,243.303333 C96.280588,240.251045 96.1618878,230.167899 96.106777,219.472176 C60.4967585,227.215235 52.9826207,204.369712 52.9826207,204.369712 C47.1599584,189.574598 38.770408,185.640538 38.770408,185.640538 C27.1568785,177.696113 39.6458206,177.859325 39.6458206,177.859325 C52.4993419,178.762293 59.267365,191.04987 59.267365,191.04987 C70.6837675,210.618423 89.2115753,204.961093 96.5158685,201.690482 C97.6647155,193.417512 100.981959,187.77078 104.642583,184.574357 C76.211799,181.33766 46.324819,170.362144 46.324819,121.315702 C46.324819,107.340889 51.3250588,95.9223682 59.5132437,86.9583937 C58.1842268,83.7344152 53.8029229,70.715562 60.7532354,53.0843636 C60.7532354,53.0843636 71.5019501,49.6441813 95.9626412,66.2049595 C106.172967,63.368876 117.123047,61.9465949 128.00106,61.8978432 C138.879073,61.9465949 149.837632,63.368876 160.067033,66.2049595 C184.49805,49.6441813 195.231926,53.0843636 195.231926,53.0843636 C202.199197,70.715562 197.815773,83.7344152 196.486756,86.9583937 C204.694018,95.9223682 209.660343,107.340889 209.660343,121.315702 C209.660343,170.478725 179.716133,181.303747 151.213281,184.472614 C155.80443,188.444828 159.895342,196.234518 159.895342,208.176593 C159.895342,225.303317 159.746968,239.087361 159.746968,243.303333 C159.746968,246.709601 162.05102,250.70089 168.53925,249.443941 C219.370432,232.499507 256,184.536204 256,128.00106 C256,57.3066942 198.691187,0 128.00106,0 Z M47.9405593,182.340212 C47.6586465,182.976105 46.6581745,183.166873 45.7467277,182.730227 C44.8183235,182.312656 44.2968914,181.445722 44.5978808,180.80771 C44.8734344,180.152739 45.876026,179.97045 46.8023103,180.409216 C47.7328342,180.826786 48.2627451,181.702199 47.9405593,182.340212 Z M54.2367892,187.958254 C53.6263318,188.524199 52.4329723,188.261363 51.6232682,187.366874 C50.7860088,186.474504 50.6291553,185.281144 51.2480912,184.70672 C51.8776254,184.140775 53.0349512,184.405731 53.8743302,185.298101 C54.7115892,186.201069 54.8748019,187.38595 54.2367892,187.958254 Z M58.5562413,195.146347 C57.7719732,195.691096 56.4895886,195.180261 55.6968417,194.042013 C54.9125733,192.903764 54.9125733,191.538713 55.713799,190.991845 C56.5086651,190.444977 57.7719732,190.936735 58.5753181,192.066505 C59.3574669,193.22383 59.3574669,194.58888 58.5562413,195.146347 Z M65.8613592,203.471174 C65.1597571,204.244846 63.6654083,204.03712 62.5716717,202.981538 C61.4524999,201.94927 61.1409122,200.484596 61.8446341,199.710926 C62.5547146,198.935137 64.0575422,199.15346 65.1597571,200.200564 C66.2704506,201.230712 66.6095936,202.705984 65.8613592,203.471174 Z M75.3025151,206.281542 C74.9930474,207.284134 73.553809,207.739857 72.1039724,207.313809 C70.6562556,206.875043 69.7087748,205.700761 70.0012857,204.687571 C70.302275,203.678621 71.7478721,203.20382 73.2083069,203.659543 C74.6539041,204.09619 75.6035048,205.261994 75.3025151,206.281542 Z M86.046947,207.473627 C86.0829806,208.529209 84.8535871,209.404622 83.3316829,209.4237 C81.8013,209.457614 80.563428,208.603398 80.5464708,207.564772 C80.5464708,206.498591 81.7483088,205.631657 83.2786917,205.606221 C84.8005962,205.576546 86.046947,206.424403 86.046947,207.473627 Z M96.6021471,207.069023 C96.7844366,208.099171 95.7267341,209.156872 94.215428,209.438785 C92.7295577,209.710099 91.3539086,209.074206 91.1652603,208.052538 C90.9808515,206.996955 92.0576306,205.939253 93.5413813,205.66582 C95.054807,205.402984 96.4092596,206.021919 96.6021471,207.069023 Z" fill="#161614"></path></g></svg></div><div class="notion-external-description"><div class="notion-external-title">download_TPAMI20_HDR.sh</div><div class="notion-external-subtitle"><span>ercanburak</span></div></div></a></li></ul><ul class="notion-list notion-list-disc notion-block-033b6e51b0df492ea468a5ae5351f629"><li>Event based: HDR from BIT，下载链接：</li></ul><ul class="notion-list notion-list-disc notion-block-56eeec28ea4b43d695059458b7890118"><li>Event guided: HDR from AKF，下载链接：<a target="_blank" rel="noopener noreferrer" href="https://github.com/ziweiWWANG/AKF" class="notion-external notion-external-mention"><div class="notion-external-image"><svg viewBox="0 0 260 260"><g><path d="M128.00106,0 C57.3172926,0 0,57.3066942 0,128.00106 C0,184.555281 36.6761997,232.535542 87.534937,249.460899 C93.9320223,250.645779 96.280588,246.684165 96.280588,243.303333 C96.280588,240.251045 96.1618878,230.167899 96.106777,219.472176 C60.4967585,227.215235 52.9826207,204.369712 52.9826207,204.369712 C47.1599584,189.574598 38.770408,185.640538 38.770408,185.640538 C27.1568785,177.696113 39.6458206,177.859325 39.6458206,177.859325 C52.4993419,178.762293 59.267365,191.04987 59.267365,191.04987 C70.6837675,210.618423 89.2115753,204.961093 96.5158685,201.690482 C97.6647155,193.417512 100.981959,187.77078 104.642583,184.574357 C76.211799,181.33766 46.324819,170.362144 46.324819,121.315702 C46.324819,107.340889 51.3250588,95.9223682 59.5132437,86.9583937 C58.1842268,83.7344152 53.8029229,70.715562 60.7532354,53.0843636 C60.7532354,53.0843636 71.5019501,49.6441813 95.9626412,66.2049595 C106.172967,63.368876 117.123047,61.9465949 128.00106,61.8978432 C138.879073,61.9465949 149.837632,63.368876 160.067033,66.2049595 C184.49805,49.6441813 195.231926,53.0843636 195.231926,53.0843636 C202.199197,70.715562 197.815773,83.7344152 196.486756,86.9583937 C204.694018,95.9223682 209.660343,107.340889 209.660343,121.315702 C209.660343,170.478725 179.716133,181.303747 151.213281,184.472614 C155.80443,188.444828 159.895342,196.234518 159.895342,208.176593 C159.895342,225.303317 159.746968,239.087361 159.746968,243.303333 C159.746968,246.709601 162.05102,250.70089 168.53925,249.443941 C219.370432,232.499507 256,184.536204 256,128.00106 C256,57.3066942 198.691187,0 128.00106,0 Z M47.9405593,182.340212 C47.6586465,182.976105 46.6581745,183.166873 45.7467277,182.730227 C44.8183235,182.312656 44.2968914,181.445722 44.5978808,180.80771 C44.8734344,180.152739 45.876026,179.97045 46.8023103,180.409216 C47.7328342,180.826786 48.2627451,181.702199 47.9405593,182.340212 Z M54.2367892,187.958254 C53.6263318,188.524199 52.4329723,188.261363 51.6232682,187.366874 C50.7860088,186.474504 50.6291553,185.281144 51.2480912,184.70672 C51.8776254,184.140775 53.0349512,184.405731 53.8743302,185.298101 C54.7115892,186.201069 54.8748019,187.38595 54.2367892,187.958254 Z M58.5562413,195.146347 C57.7719732,195.691096 56.4895886,195.180261 55.6968417,194.042013 C54.9125733,192.903764 54.9125733,191.538713 55.713799,190.991845 C56.5086651,190.444977 57.7719732,190.936735 58.5753181,192.066505 C59.3574669,193.22383 59.3574669,194.58888 58.5562413,195.146347 Z M65.8613592,203.471174 C65.1597571,204.244846 63.6654083,204.03712 62.5716717,202.981538 C61.4524999,201.94927 61.1409122,200.484596 61.8446341,199.710926 C62.5547146,198.935137 64.0575422,199.15346 65.1597571,200.200564 C66.2704506,201.230712 66.6095936,202.705984 65.8613592,203.471174 Z M75.3025151,206.281542 C74.9930474,207.284134 73.553809,207.739857 72.1039724,207.313809 C70.6562556,206.875043 69.7087748,205.700761 70.0012857,204.687571 C70.302275,203.678621 71.7478721,203.20382 73.2083069,203.659543 C74.6539041,204.09619 75.6035048,205.261994 75.3025151,206.281542 Z M86.046947,207.473627 C86.0829806,208.529209 84.8535871,209.404622 83.3316829,209.4237 C81.8013,209.457614 80.563428,208.603398 80.5464708,207.564772 C80.5464708,206.498591 81.7483088,205.631657 83.2786917,205.606221 C84.8005962,205.576546 86.046947,206.424403 86.046947,207.473627 Z M96.6021471,207.069023 C96.7844366,208.099171 95.7267341,209.156872 94.215428,209.438785 C92.7295577,209.710099 91.3539086,209.074206 91.1652603,208.052538 C90.9808515,206.996955 92.0576306,205.939253 93.5413813,205.66582 C95.054807,205.402984 96.4092596,206.021919 96.6021471,207.069023 Z" fill="#161614"></path></g></svg></div><div class="notion-external-description"><div class="notion-external-title">AKF</div><div class="notion-external-subtitle"><span>ziweiWWANG</span><span> • </span><span>Updated <!-- -->Jan 18, 2025</span></div></div></a>，<a target="_blank" rel="noopener noreferrer" href="https://github.com/ziweiWWANG/Event-Asynchronous-Filter" class="notion-external notion-external-mention"><div class="notion-external-image"><svg viewBox="0 0 260 260"><g><path d="M128.00106,0 C57.3172926,0 0,57.3066942 0,128.00106 C0,184.555281 36.6761997,232.535542 87.534937,249.460899 C93.9320223,250.645779 96.280588,246.684165 96.280588,243.303333 C96.280588,240.251045 96.1618878,230.167899 96.106777,219.472176 C60.4967585,227.215235 52.9826207,204.369712 52.9826207,204.369712 C47.1599584,189.574598 38.770408,185.640538 38.770408,185.640538 C27.1568785,177.696113 39.6458206,177.859325 39.6458206,177.859325 C52.4993419,178.762293 59.267365,191.04987 59.267365,191.04987 C70.6837675,210.618423 89.2115753,204.961093 96.5158685,201.690482 C97.6647155,193.417512 100.981959,187.77078 104.642583,184.574357 C76.211799,181.33766 46.324819,170.362144 46.324819,121.315702 C46.324819,107.340889 51.3250588,95.9223682 59.5132437,86.9583937 C58.1842268,83.7344152 53.8029229,70.715562 60.7532354,53.0843636 C60.7532354,53.0843636 71.5019501,49.6441813 95.9626412,66.2049595 C106.172967,63.368876 117.123047,61.9465949 128.00106,61.8978432 C138.879073,61.9465949 149.837632,63.368876 160.067033,66.2049595 C184.49805,49.6441813 195.231926,53.0843636 195.231926,53.0843636 C202.199197,70.715562 197.815773,83.7344152 196.486756,86.9583937 C204.694018,95.9223682 209.660343,107.340889 209.660343,121.315702 C209.660343,170.478725 179.716133,181.303747 151.213281,184.472614 C155.80443,188.444828 159.895342,196.234518 159.895342,208.176593 C159.895342,225.303317 159.746968,239.087361 159.746968,243.303333 C159.746968,246.709601 162.05102,250.70089 168.53925,249.443941 C219.370432,232.499507 256,184.536204 256,128.00106 C256,57.3066942 198.691187,0 128.00106,0 Z M47.9405593,182.340212 C47.6586465,182.976105 46.6581745,183.166873 45.7467277,182.730227 C44.8183235,182.312656 44.2968914,181.445722 44.5978808,180.80771 C44.8734344,180.152739 45.876026,179.97045 46.8023103,180.409216 C47.7328342,180.826786 48.2627451,181.702199 47.9405593,182.340212 Z M54.2367892,187.958254 C53.6263318,188.524199 52.4329723,188.261363 51.6232682,187.366874 C50.7860088,186.474504 50.6291553,185.281144 51.2480912,184.70672 C51.8776254,184.140775 53.0349512,184.405731 53.8743302,185.298101 C54.7115892,186.201069 54.8748019,187.38595 54.2367892,187.958254 Z M58.5562413,195.146347 C57.7719732,195.691096 56.4895886,195.180261 55.6968417,194.042013 C54.9125733,192.903764 54.9125733,191.538713 55.713799,190.991845 C56.5086651,190.444977 57.7719732,190.936735 58.5753181,192.066505 C59.3574669,193.22383 59.3574669,194.58888 58.5562413,195.146347 Z M65.8613592,203.471174 C65.1597571,204.244846 63.6654083,204.03712 62.5716717,202.981538 C61.4524999,201.94927 61.1409122,200.484596 61.8446341,199.710926 C62.5547146,198.935137 64.0575422,199.15346 65.1597571,200.200564 C66.2704506,201.230712 66.6095936,202.705984 65.8613592,203.471174 Z M75.3025151,206.281542 C74.9930474,207.284134 73.553809,207.739857 72.1039724,207.313809 C70.6562556,206.875043 69.7087748,205.700761 70.0012857,204.687571 C70.302275,203.678621 71.7478721,203.20382 73.2083069,203.659543 C74.6539041,204.09619 75.6035048,205.261994 75.3025151,206.281542 Z M86.046947,207.473627 C86.0829806,208.529209 84.8535871,209.404622 83.3316829,209.4237 C81.8013,209.457614 80.563428,208.603398 80.5464708,207.564772 C80.5464708,206.498591 81.7483088,205.631657 83.2786917,205.606221 C84.8005962,205.576546 86.046947,206.424403 86.046947,207.473627 Z M96.6021471,207.069023 C96.7844366,208.099171 95.7267341,209.156872 94.215428,209.438785 C92.7295577,209.710099 91.3539086,209.074206 91.1652603,208.052538 C90.9808515,206.996955 92.0576306,205.939253 93.5413813,205.66582 C95.054807,205.402984 96.4092596,206.021919 96.6021471,207.069023 Z" fill="#161614"></path></g></svg></div><div class="notion-external-description"><div class="notion-external-title">Event-Asynchronous-Filter</div><div class="notion-external-subtitle"><span>ziweiWWANG</span><span> • </span><span>Updated <!-- -->Jan 19, 2025</span></div></div></a></li></ul><ul class="notion-list notion-list-disc notion-block-96674216f25149499e25277fb83e6c2c"><li>Event guided: HDR from PKU，项目链接：<!-- -->‣<!-- -->，数据暂未公开。</li></ul><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-761de8d2589f468d811662301b6f2319" data-id="761de8d2589f468d811662301b6f2319"><span><div id="761de8d2589f468d811662301b6f2319" class="notion-header-anchor"></div><a class="notion-hash-link" href="#761de8d2589f468d811662301b6f2319" title="推荐文献"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">推荐文献</span></span></h4><div class="notion-text notion-block-1e7bd8649f0d40ba8bfc80370ca81fd9">[1] Learning to Reconstruct High Speed and High Dynamic Range Videos from Events, CVPR 2021</div><div class="notion-text notion-block-754fe4a566d3473493e6700d8632f7c3">[2] High Frame Rate Video Reconstruction Based on an Event Camera, TPAMI 2022</div><div class="notion-text notion-block-4a20c61182834776856f915981b6c0f7">[3] An Asynchronous Kalman Filter for Hybrid Event Cameras, ICCV2021</div><div class="notion-text notion-block-67458c1b3fff4b68bb9e60b8c6527ecf">[4] An Asynchronous Linear Filter Architecture for Hybrid Event-Frame Cameras, TPAMI 2023</div><div class="notion-text notion-block-fabead675ad34f2db9df1ea367781cfb">[5] Learning Event Guided High Dynamic Range Video Reconstruction, CVPR 2023</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-39d20840e1ee4972918dd2fa730ec078" data-id="39d20840e1ee4972918dd2fa730ec078"><span><div id="39d20840e1ee4972918dd2fa730ec078" class="notion-header-anchor"></div><a class="notion-hash-link" href="#39d20840e1ee4972918dd2fa730ec078" title="Video Restoration"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Video Restoration</span></span></h3><div class="notion-text notion-block-92531655805d4a9bbb6f6139f4f7da6c">由于事件数据对运动鲁棒，所以Event guided方法在运动模糊导致的质量复原任务中有奇效。</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-36397d4c90a746be85568e93f57b7bab" data-id="36397d4c90a746be85568e93f57b7bab"><span><div id="36397d4c90a746be85568e93f57b7bab" class="notion-header-anchor"></div><a class="notion-hash-link" href="#36397d4c90a746be85568e93f57b7bab" title="模拟数据集"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">模拟数据集</span></span></h4><div class="notion-text notion-block-6777b7a536394b93a3f2dea9a484e632">所谓模拟数据集，即事件数据由图像/视频模拟生成，而非真实采集得到。</div><ul class="notion-list notion-list-disc notion-block-4e34ae05c16443f089d8530ff162437a"><li>Derain，下载链接：<a target="_blank" rel="noopener noreferrer" href="https://github.com/hotndy/SPAC-SupplementaryMaterials" class="notion-external notion-external-mention"><div class="notion-external-image"><svg viewBox="0 0 260 260"><g><path d="M128.00106,0 C57.3172926,0 0,57.3066942 0,128.00106 C0,184.555281 36.6761997,232.535542 87.534937,249.460899 C93.9320223,250.645779 96.280588,246.684165 96.280588,243.303333 C96.280588,240.251045 96.1618878,230.167899 96.106777,219.472176 C60.4967585,227.215235 52.9826207,204.369712 52.9826207,204.369712 C47.1599584,189.574598 38.770408,185.640538 38.770408,185.640538 C27.1568785,177.696113 39.6458206,177.859325 39.6458206,177.859325 C52.4993419,178.762293 59.267365,191.04987 59.267365,191.04987 C70.6837675,210.618423 89.2115753,204.961093 96.5158685,201.690482 C97.6647155,193.417512 100.981959,187.77078 104.642583,184.574357 C76.211799,181.33766 46.324819,170.362144 46.324819,121.315702 C46.324819,107.340889 51.3250588,95.9223682 59.5132437,86.9583937 C58.1842268,83.7344152 53.8029229,70.715562 60.7532354,53.0843636 C60.7532354,53.0843636 71.5019501,49.6441813 95.9626412,66.2049595 C106.172967,63.368876 117.123047,61.9465949 128.00106,61.8978432 C138.879073,61.9465949 149.837632,63.368876 160.067033,66.2049595 C184.49805,49.6441813 195.231926,53.0843636 195.231926,53.0843636 C202.199197,70.715562 197.815773,83.7344152 196.486756,86.9583937 C204.694018,95.9223682 209.660343,107.340889 209.660343,121.315702 C209.660343,170.478725 179.716133,181.303747 151.213281,184.472614 C155.80443,188.444828 159.895342,196.234518 159.895342,208.176593 C159.895342,225.303317 159.746968,239.087361 159.746968,243.303333 C159.746968,246.709601 162.05102,250.70089 168.53925,249.443941 C219.370432,232.499507 256,184.536204 256,128.00106 C256,57.3066942 198.691187,0 128.00106,0 Z M47.9405593,182.340212 C47.6586465,182.976105 46.6581745,183.166873 45.7467277,182.730227 C44.8183235,182.312656 44.2968914,181.445722 44.5978808,180.80771 C44.8734344,180.152739 45.876026,179.97045 46.8023103,180.409216 C47.7328342,180.826786 48.2627451,181.702199 47.9405593,182.340212 Z M54.2367892,187.958254 C53.6263318,188.524199 52.4329723,188.261363 51.6232682,187.366874 C50.7860088,186.474504 50.6291553,185.281144 51.2480912,184.70672 C51.8776254,184.140775 53.0349512,184.405731 53.8743302,185.298101 C54.7115892,186.201069 54.8748019,187.38595 54.2367892,187.958254 Z M58.5562413,195.146347 C57.7719732,195.691096 56.4895886,195.180261 55.6968417,194.042013 C54.9125733,192.903764 54.9125733,191.538713 55.713799,190.991845 C56.5086651,190.444977 57.7719732,190.936735 58.5753181,192.066505 C59.3574669,193.22383 59.3574669,194.58888 58.5562413,195.146347 Z M65.8613592,203.471174 C65.1597571,204.244846 63.6654083,204.03712 62.5716717,202.981538 C61.4524999,201.94927 61.1409122,200.484596 61.8446341,199.710926 C62.5547146,198.935137 64.0575422,199.15346 65.1597571,200.200564 C66.2704506,201.230712 66.6095936,202.705984 65.8613592,203.471174 Z M75.3025151,206.281542 C74.9930474,207.284134 73.553809,207.739857 72.1039724,207.313809 C70.6562556,206.875043 69.7087748,205.700761 70.0012857,204.687571 C70.302275,203.678621 71.7478721,203.20382 73.2083069,203.659543 C74.6539041,204.09619 75.6035048,205.261994 75.3025151,206.281542 Z M86.046947,207.473627 C86.0829806,208.529209 84.8535871,209.404622 83.3316829,209.4237 C81.8013,209.457614 80.563428,208.603398 80.5464708,207.564772 C80.5464708,206.498591 81.7483088,205.631657 83.2786917,205.606221 C84.8005962,205.576546 86.046947,206.424403 86.046947,207.473627 Z M96.6021471,207.069023 C96.7844366,208.099171 95.7267341,209.156872 94.215428,209.438785 C92.7295577,209.710099 91.3539086,209.074206 91.1652603,208.052538 C90.9808515,206.996955 92.0576306,205.939253 93.5413813,205.66582 C95.054807,205.402984 96.4092596,206.021919 96.6021471,207.069023 Z" fill="#161614"></path></g></svg></div><div class="notion-external-description"><div class="notion-external-title">SPAC-SupplementaryMaterials</div><div class="notion-external-subtitle"><span>hotndy</span><span> • </span><span>Updated <!-- -->Oct 8, 2024</span></div></div></a></li></ul><ul class="notion-list notion-list-disc notion-block-c4cb7fa74548425fbd4cf651308d691d"><li>Deblur，下载链接：<!-- -->‣</li></ul><div class="notion-text notion-block-1b98868bf6d54d779c22994bbacfa2f0">数据处理方法：<a target="_blank" rel="noopener noreferrer" href="https://github.com/Chengzhi-Cao/SC-Net" class="notion-external notion-external-mention"><div class="notion-external-image"><svg viewBox="0 0 260 260"><g><path d="M128.00106,0 C57.3172926,0 0,57.3066942 0,128.00106 C0,184.555281 36.6761997,232.535542 87.534937,249.460899 C93.9320223,250.645779 96.280588,246.684165 96.280588,243.303333 C96.280588,240.251045 96.1618878,230.167899 96.106777,219.472176 C60.4967585,227.215235 52.9826207,204.369712 52.9826207,204.369712 C47.1599584,189.574598 38.770408,185.640538 38.770408,185.640538 C27.1568785,177.696113 39.6458206,177.859325 39.6458206,177.859325 C52.4993419,178.762293 59.267365,191.04987 59.267365,191.04987 C70.6837675,210.618423 89.2115753,204.961093 96.5158685,201.690482 C97.6647155,193.417512 100.981959,187.77078 104.642583,184.574357 C76.211799,181.33766 46.324819,170.362144 46.324819,121.315702 C46.324819,107.340889 51.3250588,95.9223682 59.5132437,86.9583937 C58.1842268,83.7344152 53.8029229,70.715562 60.7532354,53.0843636 C60.7532354,53.0843636 71.5019501,49.6441813 95.9626412,66.2049595 C106.172967,63.368876 117.123047,61.9465949 128.00106,61.8978432 C138.879073,61.9465949 149.837632,63.368876 160.067033,66.2049595 C184.49805,49.6441813 195.231926,53.0843636 195.231926,53.0843636 C202.199197,70.715562 197.815773,83.7344152 196.486756,86.9583937 C204.694018,95.9223682 209.660343,107.340889 209.660343,121.315702 C209.660343,170.478725 179.716133,181.303747 151.213281,184.472614 C155.80443,188.444828 159.895342,196.234518 159.895342,208.176593 C159.895342,225.303317 159.746968,239.087361 159.746968,243.303333 C159.746968,246.709601 162.05102,250.70089 168.53925,249.443941 C219.370432,232.499507 256,184.536204 256,128.00106 C256,57.3066942 198.691187,0 128.00106,0 Z M47.9405593,182.340212 C47.6586465,182.976105 46.6581745,183.166873 45.7467277,182.730227 C44.8183235,182.312656 44.2968914,181.445722 44.5978808,180.80771 C44.8734344,180.152739 45.876026,179.97045 46.8023103,180.409216 C47.7328342,180.826786 48.2627451,181.702199 47.9405593,182.340212 Z M54.2367892,187.958254 C53.6263318,188.524199 52.4329723,188.261363 51.6232682,187.366874 C50.7860088,186.474504 50.6291553,185.281144 51.2480912,184.70672 C51.8776254,184.140775 53.0349512,184.405731 53.8743302,185.298101 C54.7115892,186.201069 54.8748019,187.38595 54.2367892,187.958254 Z M58.5562413,195.146347 C57.7719732,195.691096 56.4895886,195.180261 55.6968417,194.042013 C54.9125733,192.903764 54.9125733,191.538713 55.713799,190.991845 C56.5086651,190.444977 57.7719732,190.936735 58.5753181,192.066505 C59.3574669,193.22383 59.3574669,194.58888 58.5562413,195.146347 Z M65.8613592,203.471174 C65.1597571,204.244846 63.6654083,204.03712 62.5716717,202.981538 C61.4524999,201.94927 61.1409122,200.484596 61.8446341,199.710926 C62.5547146,198.935137 64.0575422,199.15346 65.1597571,200.200564 C66.2704506,201.230712 66.6095936,202.705984 65.8613592,203.471174 Z M75.3025151,206.281542 C74.9930474,207.284134 73.553809,207.739857 72.1039724,207.313809 C70.6562556,206.875043 69.7087748,205.700761 70.0012857,204.687571 C70.302275,203.678621 71.7478721,203.20382 73.2083069,203.659543 C74.6539041,204.09619 75.6035048,205.261994 75.3025151,206.281542 Z M86.046947,207.473627 C86.0829806,208.529209 84.8535871,209.404622 83.3316829,209.4237 C81.8013,209.457614 80.563428,208.603398 80.5464708,207.564772 C80.5464708,206.498591 81.7483088,205.631657 83.2786917,205.606221 C84.8005962,205.576546 86.046947,206.424403 86.046947,207.473627 Z M96.6021471,207.069023 C96.7844366,208.099171 95.7267341,209.156872 94.215428,209.438785 C92.7295577,209.710099 91.3539086,209.074206 91.1652603,208.052538 C90.9808515,206.996955 92.0576306,205.939253 93.5413813,205.66582 C95.054807,205.402984 96.4092596,206.021919 96.6021471,207.069023 Z" fill="#161614"></path></g></svg></div><div class="notion-external-description"><div class="notion-external-title">SC-Net</div><div class="notion-external-subtitle"><span>Chengzhi-Cao</span><span> • </span><span>Updated <!-- -->May 22, 2024</span></div></div></a>、<a target="_blank" rel="noopener noreferrer" href="https://github.com/Chengzhi-Cao/STRA" class="notion-external notion-external-mention"><div class="notion-external-image"><svg viewBox="0 0 260 260"><g><path d="M128.00106,0 C57.3172926,0 0,57.3066942 0,128.00106 C0,184.555281 36.6761997,232.535542 87.534937,249.460899 C93.9320223,250.645779 96.280588,246.684165 96.280588,243.303333 C96.280588,240.251045 96.1618878,230.167899 96.106777,219.472176 C60.4967585,227.215235 52.9826207,204.369712 52.9826207,204.369712 C47.1599584,189.574598 38.770408,185.640538 38.770408,185.640538 C27.1568785,177.696113 39.6458206,177.859325 39.6458206,177.859325 C52.4993419,178.762293 59.267365,191.04987 59.267365,191.04987 C70.6837675,210.618423 89.2115753,204.961093 96.5158685,201.690482 C97.6647155,193.417512 100.981959,187.77078 104.642583,184.574357 C76.211799,181.33766 46.324819,170.362144 46.324819,121.315702 C46.324819,107.340889 51.3250588,95.9223682 59.5132437,86.9583937 C58.1842268,83.7344152 53.8029229,70.715562 60.7532354,53.0843636 C60.7532354,53.0843636 71.5019501,49.6441813 95.9626412,66.2049595 C106.172967,63.368876 117.123047,61.9465949 128.00106,61.8978432 C138.879073,61.9465949 149.837632,63.368876 160.067033,66.2049595 C184.49805,49.6441813 195.231926,53.0843636 195.231926,53.0843636 C202.199197,70.715562 197.815773,83.7344152 196.486756,86.9583937 C204.694018,95.9223682 209.660343,107.340889 209.660343,121.315702 C209.660343,170.478725 179.716133,181.303747 151.213281,184.472614 C155.80443,188.444828 159.895342,196.234518 159.895342,208.176593 C159.895342,225.303317 159.746968,239.087361 159.746968,243.303333 C159.746968,246.709601 162.05102,250.70089 168.53925,249.443941 C219.370432,232.499507 256,184.536204 256,128.00106 C256,57.3066942 198.691187,0 128.00106,0 Z M47.9405593,182.340212 C47.6586465,182.976105 46.6581745,183.166873 45.7467277,182.730227 C44.8183235,182.312656 44.2968914,181.445722 44.5978808,180.80771 C44.8734344,180.152739 45.876026,179.97045 46.8023103,180.409216 C47.7328342,180.826786 48.2627451,181.702199 47.9405593,182.340212 Z M54.2367892,187.958254 C53.6263318,188.524199 52.4329723,188.261363 51.6232682,187.366874 C50.7860088,186.474504 50.6291553,185.281144 51.2480912,184.70672 C51.8776254,184.140775 53.0349512,184.405731 53.8743302,185.298101 C54.7115892,186.201069 54.8748019,187.38595 54.2367892,187.958254 Z M58.5562413,195.146347 C57.7719732,195.691096 56.4895886,195.180261 55.6968417,194.042013 C54.9125733,192.903764 54.9125733,191.538713 55.713799,190.991845 C56.5086651,190.444977 57.7719732,190.936735 58.5753181,192.066505 C59.3574669,193.22383 59.3574669,194.58888 58.5562413,195.146347 Z M65.8613592,203.471174 C65.1597571,204.244846 63.6654083,204.03712 62.5716717,202.981538 C61.4524999,201.94927 61.1409122,200.484596 61.8446341,199.710926 C62.5547146,198.935137 64.0575422,199.15346 65.1597571,200.200564 C66.2704506,201.230712 66.6095936,202.705984 65.8613592,203.471174 Z M75.3025151,206.281542 C74.9930474,207.284134 73.553809,207.739857 72.1039724,207.313809 C70.6562556,206.875043 69.7087748,205.700761 70.0012857,204.687571 C70.302275,203.678621 71.7478721,203.20382 73.2083069,203.659543 C74.6539041,204.09619 75.6035048,205.261994 75.3025151,206.281542 Z M86.046947,207.473627 C86.0829806,208.529209 84.8535871,209.404622 83.3316829,209.4237 C81.8013,209.457614 80.563428,208.603398 80.5464708,207.564772 C80.5464708,206.498591 81.7483088,205.631657 83.2786917,205.606221 C84.8005962,205.576546 86.046947,206.424403 86.046947,207.473627 Z M96.6021471,207.069023 C96.7844366,208.099171 95.7267341,209.156872 94.215428,209.438785 C92.7295577,209.710099 91.3539086,209.074206 91.1652603,208.052538 C90.9808515,206.996955 92.0576306,205.939253 93.5413813,205.66582 C95.054807,205.402984 96.4092596,206.021919 96.6021471,207.069023 Z" fill="#161614"></path></g></svg></div><div class="notion-external-description"><div class="notion-external-title">STRA</div><div class="notion-external-subtitle"><span>Chengzhi-Cao</span><span> • </span><span>Updated <!-- -->Feb 13, 2024</span></div></div></a></div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-4c82598487384889999823a5e92f35d5" data-id="4c82598487384889999823a5e92f35d5"><span><div id="4c82598487384889999823a5e92f35d5" class="notion-header-anchor"></div><a class="notion-hash-link" href="#4c82598487384889999823a5e92f35d5" title="真实世界数据集"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">真实世界数据集</span></span></h4><ul class="notion-list notion-list-disc notion-block-cc5dc9d6372f4b9f988a4bb1228a1190"><li>Deblur UEVD，下载链接：<a target="_blank" rel="noopener noreferrer" href="https://github.com/intelpro/UEVD_public" class="notion-external notion-external-mention"><div class="notion-external-image"><svg viewBox="0 0 260 260"><g><path d="M128.00106,0 C57.3172926,0 0,57.3066942 0,128.00106 C0,184.555281 36.6761997,232.535542 87.534937,249.460899 C93.9320223,250.645779 96.280588,246.684165 96.280588,243.303333 C96.280588,240.251045 96.1618878,230.167899 96.106777,219.472176 C60.4967585,227.215235 52.9826207,204.369712 52.9826207,204.369712 C47.1599584,189.574598 38.770408,185.640538 38.770408,185.640538 C27.1568785,177.696113 39.6458206,177.859325 39.6458206,177.859325 C52.4993419,178.762293 59.267365,191.04987 59.267365,191.04987 C70.6837675,210.618423 89.2115753,204.961093 96.5158685,201.690482 C97.6647155,193.417512 100.981959,187.77078 104.642583,184.574357 C76.211799,181.33766 46.324819,170.362144 46.324819,121.315702 C46.324819,107.340889 51.3250588,95.9223682 59.5132437,86.9583937 C58.1842268,83.7344152 53.8029229,70.715562 60.7532354,53.0843636 C60.7532354,53.0843636 71.5019501,49.6441813 95.9626412,66.2049595 C106.172967,63.368876 117.123047,61.9465949 128.00106,61.8978432 C138.879073,61.9465949 149.837632,63.368876 160.067033,66.2049595 C184.49805,49.6441813 195.231926,53.0843636 195.231926,53.0843636 C202.199197,70.715562 197.815773,83.7344152 196.486756,86.9583937 C204.694018,95.9223682 209.660343,107.340889 209.660343,121.315702 C209.660343,170.478725 179.716133,181.303747 151.213281,184.472614 C155.80443,188.444828 159.895342,196.234518 159.895342,208.176593 C159.895342,225.303317 159.746968,239.087361 159.746968,243.303333 C159.746968,246.709601 162.05102,250.70089 168.53925,249.443941 C219.370432,232.499507 256,184.536204 256,128.00106 C256,57.3066942 198.691187,0 128.00106,0 Z M47.9405593,182.340212 C47.6586465,182.976105 46.6581745,183.166873 45.7467277,182.730227 C44.8183235,182.312656 44.2968914,181.445722 44.5978808,180.80771 C44.8734344,180.152739 45.876026,179.97045 46.8023103,180.409216 C47.7328342,180.826786 48.2627451,181.702199 47.9405593,182.340212 Z M54.2367892,187.958254 C53.6263318,188.524199 52.4329723,188.261363 51.6232682,187.366874 C50.7860088,186.474504 50.6291553,185.281144 51.2480912,184.70672 C51.8776254,184.140775 53.0349512,184.405731 53.8743302,185.298101 C54.7115892,186.201069 54.8748019,187.38595 54.2367892,187.958254 Z M58.5562413,195.146347 C57.7719732,195.691096 56.4895886,195.180261 55.6968417,194.042013 C54.9125733,192.903764 54.9125733,191.538713 55.713799,190.991845 C56.5086651,190.444977 57.7719732,190.936735 58.5753181,192.066505 C59.3574669,193.22383 59.3574669,194.58888 58.5562413,195.146347 Z M65.8613592,203.471174 C65.1597571,204.244846 63.6654083,204.03712 62.5716717,202.981538 C61.4524999,201.94927 61.1409122,200.484596 61.8446341,199.710926 C62.5547146,198.935137 64.0575422,199.15346 65.1597571,200.200564 C66.2704506,201.230712 66.6095936,202.705984 65.8613592,203.471174 Z M75.3025151,206.281542 C74.9930474,207.284134 73.553809,207.739857 72.1039724,207.313809 C70.6562556,206.875043 69.7087748,205.700761 70.0012857,204.687571 C70.302275,203.678621 71.7478721,203.20382 73.2083069,203.659543 C74.6539041,204.09619 75.6035048,205.261994 75.3025151,206.281542 Z M86.046947,207.473627 C86.0829806,208.529209 84.8535871,209.404622 83.3316829,209.4237 C81.8013,209.457614 80.563428,208.603398 80.5464708,207.564772 C80.5464708,206.498591 81.7483088,205.631657 83.2786917,205.606221 C84.8005962,205.576546 86.046947,206.424403 86.046947,207.473627 Z M96.6021471,207.069023 C96.7844366,208.099171 95.7267341,209.156872 94.215428,209.438785 C92.7295577,209.710099 91.3539086,209.074206 91.1652603,208.052538 C90.9808515,206.996955 92.0576306,205.939253 93.5413813,205.66582 C95.054807,205.402984 96.4092596,206.021919 96.6021471,207.069023 Z" fill="#161614"></path></g></svg></div><div class="notion-external-description"><div class="notion-external-title">UEVD_public</div><div class="notion-external-subtitle"><span>intelpro</span><span> • </span><span>Updated <!-- -->Nov 12, 2024</span></div></div></a></li></ul><ul class="notion-list notion-list-disc notion-block-05b0c94ec9a94eaca0eaf6954594fd92"><li>Deblur FEVD，下载链接：<!-- -->‣</li></ul><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-53624ab6ef204f89991dd55f9cc73bf9" data-id="53624ab6ef204f89991dd55f9cc73bf9"><span><div id="53624ab6ef204f89991dd55f9cc73bf9" class="notion-header-anchor"></div><a class="notion-hash-link" href="#53624ab6ef204f89991dd55f9cc73bf9" title="推荐文献"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">推荐文献</span></span></h4><div class="notion-text notion-block-fec8bbac6092486393e29dd02ef7c1e5">[1] Bringing Events Into Video Deblurring With Non-Consecutively Blurry Frames, ICCV 2021</div><div class="notion-text notion-block-73daa162526b43b2b4ce6305c44adcb0">[2] Event-driven Video Deblurring via Spatio-Temporal Relation-Aware Network, IJCAI 2022</div><div class="notion-text notion-block-4ee5492e98284ec494f75820bf8e4a15">[3] Event-guided Deblurring of Unknown Exposure Time Videos, ECCV 2022</div><div class="notion-text notion-block-e051596d5d684579a6d4ecec2da4009a">[4] Event-Driven Video Restoration With Spiking-Convolutional Architecture, TNNLS 2023</div><div class="notion-text notion-block-235c1fd5e47e4881b4af4399652ffa25">[5] Frequency-aware Event-based Video Deblurring for Real-World Motion Blur, CVPR 2024</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-6937b272d0524099bee554d9fac3f85b" data-id="6937b272d0524099bee554d9fac3f85b"><span><div id="6937b272d0524099bee554d9fac3f85b" class="notion-header-anchor"></div><a class="notion-hash-link" href="#6937b272d0524099bee554d9fac3f85b" title="Video Enhancement"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Video Enhancement</span></span></h3><div class="notion-text notion-block-69994b9d9ee34f22858b7d160ff2c23f">由于事件数据在暗夜条件下也有不错的运动检测能力，所以Event guided方法在暗夜增强等复原任务中也有使用。</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-06aa5a17820c4964a54c3126cc06754f" data-id="06aa5a17820c4964a54c3126cc06754f"><span><div id="06aa5a17820c4964a54c3126cc06754f" class="notion-header-anchor"></div><a class="notion-hash-link" href="#06aa5a17820c4964a54c3126cc06754f" title="真实数据集"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">真实数据集</span></span></h4><ul class="notion-list notion-list-disc notion-block-97b53a580cbe41f5a2b465516c4ab53b"><li>Deblur and Low-light，下载链接：<a target="_blank" rel="noopener noreferrer" href="https://github.com/fourson/Deblurring-Low-Light-Images-with-Events" class="notion-external notion-external-mention"><div class="notion-external-image"><svg viewBox="0 0 260 260"><g><path d="M128.00106,0 C57.3172926,0 0,57.3066942 0,128.00106 C0,184.555281 36.6761997,232.535542 87.534937,249.460899 C93.9320223,250.645779 96.280588,246.684165 96.280588,243.303333 C96.280588,240.251045 96.1618878,230.167899 96.106777,219.472176 C60.4967585,227.215235 52.9826207,204.369712 52.9826207,204.369712 C47.1599584,189.574598 38.770408,185.640538 38.770408,185.640538 C27.1568785,177.696113 39.6458206,177.859325 39.6458206,177.859325 C52.4993419,178.762293 59.267365,191.04987 59.267365,191.04987 C70.6837675,210.618423 89.2115753,204.961093 96.5158685,201.690482 C97.6647155,193.417512 100.981959,187.77078 104.642583,184.574357 C76.211799,181.33766 46.324819,170.362144 46.324819,121.315702 C46.324819,107.340889 51.3250588,95.9223682 59.5132437,86.9583937 C58.1842268,83.7344152 53.8029229,70.715562 60.7532354,53.0843636 C60.7532354,53.0843636 71.5019501,49.6441813 95.9626412,66.2049595 C106.172967,63.368876 117.123047,61.9465949 128.00106,61.8978432 C138.879073,61.9465949 149.837632,63.368876 160.067033,66.2049595 C184.49805,49.6441813 195.231926,53.0843636 195.231926,53.0843636 C202.199197,70.715562 197.815773,83.7344152 196.486756,86.9583937 C204.694018,95.9223682 209.660343,107.340889 209.660343,121.315702 C209.660343,170.478725 179.716133,181.303747 151.213281,184.472614 C155.80443,188.444828 159.895342,196.234518 159.895342,208.176593 C159.895342,225.303317 159.746968,239.087361 159.746968,243.303333 C159.746968,246.709601 162.05102,250.70089 168.53925,249.443941 C219.370432,232.499507 256,184.536204 256,128.00106 C256,57.3066942 198.691187,0 128.00106,0 Z M47.9405593,182.340212 C47.6586465,182.976105 46.6581745,183.166873 45.7467277,182.730227 C44.8183235,182.312656 44.2968914,181.445722 44.5978808,180.80771 C44.8734344,180.152739 45.876026,179.97045 46.8023103,180.409216 C47.7328342,180.826786 48.2627451,181.702199 47.9405593,182.340212 Z M54.2367892,187.958254 C53.6263318,188.524199 52.4329723,188.261363 51.6232682,187.366874 C50.7860088,186.474504 50.6291553,185.281144 51.2480912,184.70672 C51.8776254,184.140775 53.0349512,184.405731 53.8743302,185.298101 C54.7115892,186.201069 54.8748019,187.38595 54.2367892,187.958254 Z M58.5562413,195.146347 C57.7719732,195.691096 56.4895886,195.180261 55.6968417,194.042013 C54.9125733,192.903764 54.9125733,191.538713 55.713799,190.991845 C56.5086651,190.444977 57.7719732,190.936735 58.5753181,192.066505 C59.3574669,193.22383 59.3574669,194.58888 58.5562413,195.146347 Z M65.8613592,203.471174 C65.1597571,204.244846 63.6654083,204.03712 62.5716717,202.981538 C61.4524999,201.94927 61.1409122,200.484596 61.8446341,199.710926 C62.5547146,198.935137 64.0575422,199.15346 65.1597571,200.200564 C66.2704506,201.230712 66.6095936,202.705984 65.8613592,203.471174 Z M75.3025151,206.281542 C74.9930474,207.284134 73.553809,207.739857 72.1039724,207.313809 C70.6562556,206.875043 69.7087748,205.700761 70.0012857,204.687571 C70.302275,203.678621 71.7478721,203.20382 73.2083069,203.659543 C74.6539041,204.09619 75.6035048,205.261994 75.3025151,206.281542 Z M86.046947,207.473627 C86.0829806,208.529209 84.8535871,209.404622 83.3316829,209.4237 C81.8013,209.457614 80.563428,208.603398 80.5464708,207.564772 C80.5464708,206.498591 81.7483088,205.631657 83.2786917,205.606221 C84.8005962,205.576546 86.046947,206.424403 86.046947,207.473627 Z M96.6021471,207.069023 C96.7844366,208.099171 95.7267341,209.156872 94.215428,209.438785 C92.7295577,209.710099 91.3539086,209.074206 91.1652603,208.052538 C90.9808515,206.996955 92.0576306,205.939253 93.5413813,205.66582 C95.054807,205.402984 96.4092596,206.021919 96.6021471,207.069023 Z" fill="#161614"></path></g></svg></div><div class="notion-external-description"><div class="notion-external-title">Deblurring-Low-Light-Images-with-Events</div><div class="notion-external-subtitle"><span>fourson</span><span> • </span><span>Updated <!-- -->Dec 1, 2024</span></div></div></a></li></ul><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-cdb0e580d82448ce979efec9d5a5ef93" data-id="cdb0e580d82448ce979efec9d5a5ef93"><span><div id="cdb0e580d82448ce979efec9d5a5ef93" class="notion-header-anchor"></div><a class="notion-hash-link" href="#cdb0e580d82448ce979efec9d5a5ef93" title="推荐文献"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">推荐文献</span></span></h4><div class="notion-text notion-block-77c29a96584749a5bbda4e6e1f8efa87">[1] Deblurring Low-Light Images with Events, IJCV 2023</div><div class="notion-text notion-block-eb79ae85ba824f9584b0a2a59c905a13">[2] Coherent Event Guided Low-Light Video Enhancement, CVPR 2023</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-c460eee03eb1485393c70e59d1fd1f9f" data-id="c460eee03eb1485393c70e59d1fd1f9f"><span><div id="c460eee03eb1485393c70e59d1fd1f9f" class="notion-header-anchor"></div><a class="notion-hash-link" href="#c460eee03eb1485393c70e59d1fd1f9f" title="Video Super Resolution"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Video Super Resolution</span></span></h3><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-6167ec766f484c0fb4db8785531a1988" data-id="6167ec766f484c0fb4db8785531a1988"><span><div id="6167ec766f484c0fb4db8785531a1988" class="notion-header-anchor"></div><a class="notion-hash-link" href="#6167ec766f484c0fb4db8785531a1988" title="Video Frame Interpolation"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Video Frame Interpolation</span></span></h3><div class="notion-blank notion-block-edfd07e1a4944b75b8caed4ef4acb0ef"> </div><h2 class="notion-h notion-h1 notion-h-indent-0 notion-block-de25877fef2b4158bd868c054da76d1c" data-id="de25877fef2b4158bd868c054da76d1c"><span><div id="de25877fef2b4158bd868c054da76d1c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#de25877fef2b4158bd868c054da76d1c" title="Event Output"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Event Output</span></span></h2><div class="notion-text notion-block-31b1d0df82bb4f70b4666df138283f16">最后值得注意的是，纯粹对事件数据进行复原也是一项任务。当然与部分医学图像复原任务类似，此类任务应当依赖成像技术的更迭而非算法技术辅助。</div><div class="notion-blank notion-block-a439e77c03cd46899df375d3a23a7d1f"> </div></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Unsupervised (unpaired) learning in blind super resolution]]></title>
            <link>https://blog.jongkhu.com//article/unsupervised-super-resolution</link>
            <guid>https://blog.jongkhu.com//article/unsupervised-super-resolution</guid>
            <pubDate>Tue, 04 Apr 2023 00:00:00 GMT</pubDate>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-full-width notion-block-7c711be298db4b159d52caad9ff0d09e"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><div class="notion-text notion-block-840971314dcd411ea997bf987f3b285a">本文简单讲述一下无监督学习在盲超分任务中的应用，主要有域适应、对比学习、模糊核估计等方法。<div class="notion-text-children"><blockquote class="notion-quote notion-block-6146991ccf5147a99da4e0ab75204465"><div>超分任务的无监督并不是很严格，很多无监督主要表现形式为：非配对，本质上还是需要HR图像的信息对网络进行监督。<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://ar5iv.labs.arxiv.org/html/2004.11020v1">SimUSR: A Simple but Strong Baseline for Unsupervised Image Super-resolution</a>等文中提出了一些不需要HR的无监督超分方法。</div></blockquote></div></div><h2 class="notion-h notion-h1 notion-h-indent-0 notion-block-a729714e0fce407082651bf49d7d4dc3" data-id="a729714e0fce407082651bf49d7d4dc3"><span><div id="a729714e0fce407082651bf49d7d4dc3" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a729714e0fce407082651bf49d7d4dc3" title="盲超分任务"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">盲超分任务</span></span></h2><div class="notion-text notion-block-e2debd1a726f440eb8421d8b213dc1c6">盲超分任务：”Blind image super-resolution (SR), aiming to super-resolve low-resolution images with unknown degradation.”</div><div class="notion-text notion-block-4c81f30c81404d76ab4148290cef9b62">其目的是解决真实世界的超分问题，一般的基于深度学习的超分往往只在简单的下采样核，例如：bicubic，的情况下使用神经网络对LR-HR图像对进行拟合，这会使得网络仅能处理几个经典的下采样核场景而无法在真实场景下完成超分。</div><blockquote class="notion-quote notion-block-69e5b3e52bad4194b609cf2e628cfbd2"><div>引自”Blind Image Super-Resolution: A Survey and Beyond”, Anran Liu, Yihao Liu, Jinjin Gu, Yu Qiao, and Chao Dong</div></blockquote><div class="notion-text notion-block-4433366ae33947b69742b06c031c203d">同时由于真实的、成对的LR-HR图像对难以获取，无监督方案越来越受到研究人员关注。</div><blockquote class="notion-quote notion-block-3a79fa2ab8bc4041900a3c349006b8ff"><div>为方便叙述，后文中所有传统下采样核均被表述为bicubic；大多数文章的实验为正规的对比实验+消融实验，这里仅挑有意思的实验讲解</div></blockquote><h2 class="notion-h notion-h1 notion-h-indent-0 notion-block-2da06ba02b3a4944a0da2675beaa436d" data-id="2da06ba02b3a4944a0da2675beaa436d"><span><div id="2da06ba02b3a4944a0da2675beaa436d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#2da06ba02b3a4944a0da2675beaa436d" title="Domain Adaptation"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Domain Adaptation</span></span></h2><blockquote class="notion-quote notion-block-2117f2d4659e4b0f81b141fa2dc8298a"><div>在<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://openaccess.thecvf.com/content/CVPR2022/papers/Xu_Dual_Adversarial_Adaptation_for_Cross-Device_Real-World_Image_Super-Resolution_CVPR_2022_paper.pdf">Cross Device Super Resolution</a>那篇文章中专门提到了UDA在超分等图像复原任务中的难点：”Few are devoted to low-level vision, which is more challenging for UDA since it concentrates more on pixel-wise adaptation and is not easy just by simple feature alignment or sample distribution alignment like UDA in high-level vision.”</div></blockquote><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-7f23fd7de71c45df9a7b9899c93104d0" data-id="7f23fd7de71c45df9a7b9899c93104d0"><span><div id="7f23fd7de71c45df9a7b9899c93104d0" class="notion-header-anchor"></div><a class="notion-hash-link" href="#7f23fd7de71c45df9a7b9899c93104d0" title="DASR"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">DASR</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-48087ec0f2d3414d956fb889414f0ec9" data-id="48087ec0f2d3414d956fb889414f0ec9"><span><div id="48087ec0f2d3414d956fb889414f0ec9" class="notion-header-anchor"></div><a class="notion-hash-link" href="#48087ec0f2d3414d956fb889414f0ec9" title="文章简介"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">文章简介</span></span></h4><div class="notion-text notion-gray_background notion-block-34ccca4d0506490a8629d8b05605821c">文章全称：Unsupervised Real-world Image Super Resolution via Domain-distance Aware Training</div><div class="notion-text notion-gray_background notion-block-a8bd8abd021f4534b4ab356a5d531597">文章链接：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://ar5iv.labs.arxiv.org/html/2004.01178v1">https://ar5iv.labs.arxiv.org/html/2004.01178v1</a></div><div class="notion-text notion-gray_background notion-block-860db90024014a319dd013ff6c78bb44">文章代码：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/ShuhangGu/DASR">https://github.com/ShuhangGu/DASR</a></div><div class="notion-text notion-block-19c6d7eac3cd4cb8ac68ee3ab24c1cb5">一般基于深度学习的超分使用bicubic制作数据，该方案无法适用于真实场景，换句话说就是bicubic制作的LR图像与真实世界的LR图像存在差异。文章作者的想法十分简洁：如果我们需要对真实图像进行超分，拉进bicubic图像与真实图像的距离就好了。</div><div class="notion-text notion-block-23652c9adb4643c8b1589c27ee843f78">笔者认为本文中心是原文abstract中的这句话：<b>”Domain-gap aware training takes additional benefit from real data in the target domain while domain-distance weighted supervision brings forward the more rational use of labeled source domain data.” </b>这句话说明了文章的两个重点：</div><ol start="1" class="notion-list notion-list-numbered notion-block-86f65cb7e71d48179a1ffb25b84b919d"><li>domain-gap aware training是有非配对真实数据作为输入的，它的目的是拉进目标domain（真实图像域）与现有domain（bicubic图像域或由网络生成的图像域）的距离</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-db54232055044e5e9819cb9968488ebd"><li>domain-distance weighted supervision能够更好的利用现有domain（bicubic图像域或由网络生成的图像域）的数据</li></ol><div class="notion-text notion-block-d4ab9d2bfc0041a18e219a8993655650">综合起来就是用一系列手段拉进了bicubic域与真实图像域的距离，能够更好的超分真实图像了。</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-a392af005d674400aa3611a408d121be" data-id="a392af005d674400aa3611a408d121be"><span><div id="a392af005d674400aa3611a408d121be" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a392af005d674400aa3611a408d121be" title="解决方案"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">解决方案</span></span></h4><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-42747d1c3ffe48bea1c3ed46b92a673c"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://ar5iv.labs.arxiv.org/html/2004.01178/assets/figures/Overall.png?t=42747d1c-3ffe-48be-a1c3-ed46b92a673c" alt="notion image" loading="lazy" decoding="async"/></div></figure><ol start="1" class="notion-list notion-list-numbered notion-block-883fc640912e4f179751e1f9f58c80f4"><li>Training of Down-Sampling Network</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-a14be9d75d77492799288da48e1e2b0f"><li>Domain distance aware training of Super-Resolution Network</li></ol><div class="notion-text notion-block-2f024ef48ca648d8a8a92724343d70ae">整个训练方案包含两部分：DSN和SRN，DSN负责从HR图像学习到真实的LR图像，SRN负责超分，接下来分别介绍这两个模块。</div><div class="notion-text notion-block-d2e76f66126945caaf18c63d841becbc">DSN本质上是一个<b>Unpaired GAN网络</b>，输入真实高分辨率图像，输入1. 真实低分辨率图像，2. domain distance map。为了避免unpaired训练导致的颜色偏移或者伪影出现，在实际训练中只把高频信息（细节部分，也就是退化更明显的部分）输入至判别器。</div><blockquote class="notion-quote notion-block-547c4fb8af144888b4c932bbef6ce93f"><div>这里想法比较像同期的一篇”Deblurring by Realistic Blurring”。同时这个高频假设比较牵强，高频≠退化更明显，文章仅从实验角度简单分析了这一假设</div></blockquote><div class="notion-text notion-block-738f4e9d0c874a40aa6c14d1ee604c46">domain distance map 表示了判别器认为的每一个patch的真实情况，值越高代表该patch越可能是真实图像。</div><div class="notion-text notion-block-c1448b76e460460e8096dda3f6688754">DDM将被嵌入到SRN中辅助超分网络的训练，即<b>在原始pixel loss基础上点乘domain distance map，使得domain distance map中值更小的部分也就是更真实的部分的loss被突出</b>，使得网络更关注真实图像的超分问题。</div><div class="notion-text notion-block-84c48718aacb4ee7abed6df470b0a011">同时为了进一步确定SRN生成的图像是真实的，作者加入了一个额外的对抗损失：</div><div class="notion-text notion-block-cc7d1ba444bf446598d04afbc49964ef">learned角标代表该图像由DSN生成，该loss的主要目的就是保证即使DSN学不到真实图像的下采样方法，也能有这个对抗损失给它兜底，让SRN学到对真实图像的上采样。</div><blockquote class="notion-quote notion-block-55c87860ee324aef81ef0464cac2c4cc"><div>这个SRN虽然存在真实的LR输入，但是是非配对的任意真实LR图像，这不破坏非配对这一前提</div></blockquote><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-bf87db18bb714bb3969e4567f742557d" data-id="bf87db18bb714bb3969e4567f742557d"><span><div id="bf87db18bb714bb3969e4567f742557d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#bf87db18bb714bb3969e4567f742557d" title="一些疑问"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">一些疑问</span></span></h4><div class="notion-text notion-block-ad665639242a4fdbbfaa254a809e301d">这篇文章想法很简洁，方法比较精妙。但是个人觉得存在以下一个问题：</div><ol start="1" class="notion-list notion-list-numbered notion-block-8977dc5188b449428cfbc2c3ea1236b8"><li>文章提取了图像的高频信息后输入DSN进入判别器，这里其实还暗含一个假设：图像高频信息代表的退化模式即为全图的退化模式，这是不合理的。</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-8803d0f5ea7d4d4f86bd67278500c182"><li>如果提取特征的高频信息呢？</li></ol><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-b3629b9fad2449578cd0958478a52511" data-id="b3629b9fad2449578cd0958478a52511"><span><div id="b3629b9fad2449578cd0958478a52511" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b3629b9fad2449578cd0958478a52511" title="USR-DA"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">USR-DA</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-aea0b4df1852469fab79177f3c377ee4" data-id="aea0b4df1852469fab79177f3c377ee4"><span><div id="aea0b4df1852469fab79177f3c377ee4" class="notion-header-anchor"></div><a class="notion-hash-link" href="#aea0b4df1852469fab79177f3c377ee4" title="文章简介"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">文章简介</span></span></h4><div class="notion-text notion-gray_background notion-block-22149f8c74724748957240d5b8e5e267">文章全称：Unsupervised Real-World Super-Resolution: A Domain Adaptation Perspective</div><div class="notion-text notion-gray_background notion-block-b31027607790494c9a9ff3163d4c571c">文章链接：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://ieeexplore.ieee.org/document/9711325">https://ieeexplore.ieee.org/document/9711325</a></div><div class="notion-text notion-gray_background notion-block-fe51a52ba03541d094c0f52817c01e90">文章代码：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/anse3832/USR_DA">https://github.com/anse3832/USR_DA</a> (unofficial)</div><div class="notion-text notion-block-5396197e28294dd8ba1c7252d456ff9c">这篇文章把拉进bicubic域与真实图像域的问题视为domain adaptation问题，文章认为由深度神经网络学习如何真实的降采样是不稳定的（two-stage方法的通病）。所以，”We propose a novel unpaired SR training framework based on <b>feature distribution alignment</b>, with which we can obtain degradation-indistinguishable feature maps and then map them to HR images.”</div><blockquote class="notion-quote notion-block-56bdea32977d438dbc883524f99151b1"><div>上一篇在图像空间拉进两个domain的距离，这篇是在特征空间拉进两个domain的距离</div></blockquote><div class="notion-text notion-block-6cec26509d1c4928a8e22a26d8b372fe">所以，文章主要要解决的问题就是：</div><ol start="1" class="notion-list notion-list-numbered notion-block-63887fba732048d3bf0c295ed037f8bf"><li><b><b>Feature Distribution Alignment：</b></b>如何将bicubic域与真实域的LR图像投影到一个特征空间中，保证网络（encoder）能够学习到与退化模型无关的特征</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-7b490f652af444cb81ab000f29d918d2"><li><b><b>Feature Domain Regularization：</b></b>使得上述特征空间能够保留更多真实图像的信息，以确保超分网络在处理真实图像的特征</li></ol><div class="notion-text notion-block-3bd32940005046acbdc0b87cfa760dd4">文章的图画的很好：</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-62ee0b5838084574b0321bebf91e1819"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://ieeexplore.ieee.org/mediastore_new/IEEE/content/media/9709627/9709628/9711325/wang1-p10-wang-large.gif?t=62ee0b58-3808-4574-b032-1bebf91e1819" alt="notion image" loading="lazy" decoding="async"/></div></figure><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-5e58e3acecf24a6b8d3eeb16a3d89eae" data-id="5e58e3acecf24a6b8d3eeb16a3d89eae"><span><div id="5e58e3acecf24a6b8d3eeb16a3d89eae" class="notion-header-anchor"></div><a class="notion-hash-link" href="#5e58e3acecf24a6b8d3eeb16a3d89eae" title="解决方案"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">解决方案</span></span></h4><div class="notion-text notion-block-6b6d76872b114d5cb062d0d3ad336645">这篇文章的解决方案比较复杂。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-2b71617b371d46aa8f8c97038a661440"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://ieeexplore.ieee.org/mediastore_new/IEEE/content/media/9709627/9709628/9711325/wang2-p10-wang-large.gif?t=2b71617b-371d-46aa-8f8c-97038a661440" alt="notion image" loading="lazy" decoding="async"/></div></figure><blockquote class="notion-quote notion-block-4d57c96db5964d4481bd7d5a4dc7f1e0"><div>粉红色路线为推理路线</div></blockquote><div class="notion-text notion-block-f597136b482046d6ad2928d16c93a0d3">与上文一样，我们已知成对的bicubic域的LR-HR图像对，和一个随机的真实域的LR图像。</div><div class="notion-text notion-block-f9e135c462bb419894d171f7079d4b3a">其中，Feature Distribution Alignment使用GAN完成，比较简单，我们主要关注Feature Domain Regularization部分。</div><div class="notion-text notion-block-1cb43c1905cb4419a84357226246035a">Feature Domain Regularization需要做到以下两个目标：</div><ol start="1" class="notion-list notion-list-numbered notion-block-1b93f4afeffb47abb0fd86980f2f7f80"><li>共享特征空间不会损失bicubic域LR图像的内容</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-49272d1012bd49c4a4180604331fdd46"><li>共享特征空间不会损失真实域LR图像的退化模型的抽象表征</li></ol><div class="notion-text notion-block-64d3862b312b431aa927d5efccc2b023">当然是用CycleGAN：</div><ul class="notion-list notion-list-disc notion-block-43b8bd732c0845b6a6d95e511930004a"><li>Gt负责把不同域图像的特征都映射到真实域中</li></ul><ul class="notion-list notion-list-disc notion-block-55ab62b6774644ab8c30c9b0b9aa508a"><li>图中右侧E负责从由Gt得到的真实域LR图像提取出共享特征空间的特征，它包含真实域的特征信息，并输入超分网络</li></ul><ul class="notion-list notion-list-disc notion-block-d9cc5ab71afe4520959272fec8af1a44"><li>超分网络存在两个输入，一个是bicubic域图像的在共享特征空间中的特征，一个是bicubic域图像经过了Gt所生成的真实域图像的在共享特征空间中的特征，分别输出对应的超分结果</li></ul><ul class="notion-list notion-list-disc notion-block-0fa231e74dea4353a60a6d9b93c0983b"><li>拉进：输入超分网络的特征（权重较小，比较陷入两个特征完全一样的局部最优）；超分网络两个输出与原始HR的距离；同时需要用GAN处理一下保证Gt输出的图像在判别器眼里是真实域的</li></ul><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-ac7c0b95a4eb4226affe735bda86ef4b" data-id="ac7c0b95a4eb4226affe735bda86ef4b"><span><div id="ac7c0b95a4eb4226affe735bda86ef4b" class="notion-header-anchor"></div><a class="notion-hash-link" href="#ac7c0b95a4eb4226affe735bda86ef4b" title="一些疑问"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">一些疑问</span></span></h4><div class="notion-text notion-block-96dc6e887fe44fcd981df0ff1176e8ae">这篇文章就是将UDA思想迁入超分任务中，并且loss很多不是很简洁。笔者对这篇文章还有一些疑问：</div><ol start="1" class="notion-list notion-list-numbered notion-block-941332fa2c5348dcbe2c728c2e896724"><li>学习网络把bicubic和真实图像投射到同一特征空间这一操作其实就是做了一个，但是这样会导致真实LR的信息消失，笔者认为这种操作不是很好，放在去噪等任务可能更有说服力。</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-69ac6f322d6040aca7aa7535374aaba9"><li>一开始把bicubic和真实图像投射到同一特征空间后又需要让该空间保留真实域的信息，这种对抗似乎比原始的GAN更难训练。</li></ol><ol start="3" class="notion-list notion-list-numbered notion-block-ef5349085b82462abddb9a83124cefba"><li>如果直接用共享空间信息与真实LR图像一起作为输入肯定比现有方案更好，这样就不需要把共享空间拉往真实域空间，但是这样就无法非配对了。这篇文章有一点为了这碟醋包饺子的嫌疑。</li></ol><h2 class="notion-h notion-h1 notion-h-indent-0 notion-block-59b6c5c5d96f4863a6f556fd1498fc0d" data-id="59b6c5c5d96f4863a6f556fd1498fc0d"><span><div id="59b6c5c5d96f4863a6f556fd1498fc0d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#59b6c5c5d96f4863a6f556fd1498fc0d" title="Contrastive Learning"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Contrastive Learning</span></span></h2><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-a76b048e043a4a3d8b935a7dae634ca2" data-id="a76b048e043a4a3d8b935a7dae634ca2"><span><div id="a76b048e043a4a3d8b935a7dae634ca2" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a76b048e043a4a3d8b935a7dae634ca2" title="DASR"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">DASR</span></span></h3><blockquote class="notion-quote notion-block-3645b1c6556145748744b9af8a5e836b"><div>又一个DASR，</div></blockquote><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-58d7214d9ce940fc88be7e6e1cfa1108" data-id="58d7214d9ce940fc88be7e6e1cfa1108"><span><div id="58d7214d9ce940fc88be7e6e1cfa1108" class="notion-header-anchor"></div><a class="notion-hash-link" href="#58d7214d9ce940fc88be7e6e1cfa1108" title="文章简介"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">文章简介</span></span></h4><div class="notion-text notion-gray_background notion-block-849b1f8373204f8db1bfdb6b9dbc1674">文章全称：Unsupervised Degradation Representation Learning for Blind Super-Resolution</div><div class="notion-text notion-gray_background notion-block-d86a582368334349939f5068420cffea">文章链接：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://ar5iv.labs.arxiv.org/html/2104.00416v1">https://ar5iv.labs.arxiv.org/html/2104.00416v1</a></div><div class="notion-text notion-gray_background notion-block-722cf98887d24d9aaa6fc868ad209893">文章代码：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/The-Learning-And-Vision-Atelier-LAVA/DASR">https://github.com/The-Learning-And-Vision-Atelier-LAVA/DASR</a></div><div class="notion-text notion-block-b69d350bafc64df38306ba9bdabd7d11">目前盲超分任务很多文章主要目的是解决特定的退化核，有些盲超分方案甚至无法有效超分bicubic下采样的图像，这与盲超分任务的初衷背道而驰。为了能够解决现实中出现的未知退化，<b>需要将退化模型从在像素空间的显式估计转移为在抽象表示，即原文abstract中的”Specifically, we learn abstract representations to distinguish various degradations in the representation space rather than explicit estimation in the pixel space.”</b></div><div class="notion-text notion-block-61ef55b302f64e149ce814b1e1d904f4">文章的introduction部分说明了该方案的两个优点：</div><ol start="1" class="notion-list notion-list-numbered notion-block-6ce83392ce014f819c21dbe2d64ca83d"><li><b>First</b>, compared to extracting full representations to estimate degradations, it is easier to learn abstract representations to distinguish different degradations.</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-26abe3f9a66a46bc90af0c812ca336fb"><li><b>Second</b>, degradation representation learning does not require the supervision from ground truth degradation.</li></ol><div class="notion-text notion-block-86db718190694e71adb1b8a8cc64297a">即，特异性 + 不需要GT。特异性指，这个方法能够使得退化估计模块能够针对不同图像得到不同的退化估计结果。</div><blockquote class="notion-quote notion-block-b9e690f5a63c46618e0adc99aa952188"><div>这一点其实有待商榷，成像结果的退化模型有很大一部分受硬件影响，在一些特殊领域可能是图像不同patch之间具有不同的退化模型。</div></blockquote><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-d4eb636d04974563ae7ee0567235518f" data-id="d4eb636d04974563ae7ee0567235518f"><span><div id="d4eb636d04974563ae7ee0567235518f" class="notion-header-anchor"></div><a class="notion-hash-link" href="#d4eb636d04974563ae7ee0567235518f" title="解决方案"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">解决方案</span></span></h4><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-e2e8e5e0ea3d469dbba7471ab2246aeb"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://ar5iv.labs.arxiv.org/html/2104.00416/assets/x2.png?t=e2e8e5e0-ea3d-469d-bba7-471ab2246aeb" alt="notion image" loading="lazy" decoding="async"/></div></figure><div class="notion-text notion-block-f17028eb6a784aa2a2e13be0df71e390"><b>we assume the degradation is the same in an image but can vary for different images</b></div><ol start="1" class="notion-list notion-list-numbered notion-block-6711395335fa40cca9c8d51042ed442c"><li>an image patch should <b>be similar to other patches in the same image</b> (<em>i.e.</em>, with the same degradation)</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-a173e72b09e446a2a23c44c4f57819a3"><li>an image patch should <b>be dissimilar to patches from other images</b> (<em>i.e.</em>, with different degradation)</li></ol><div class="notion-text notion-block-e11e2dd9962c4c3f818825c29eed69c6">文章通过拉进同一图像不同patch之间的退化估计结果、拉远不同图像patch之间的退化估计结果，进行对比学习。</div><div class="notion-text notion-block-e3f89b0cdf104ffbb7a3261ed01a8b6e">把退化模型的抽象表述嵌入超分网络的方案类似于SFT，具体的特征整合方案烦请读者自行查看代码，在此不再赘述。</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-60617a5df4664afdb9d52629c6a4c2ac" data-id="60617a5df4664afdb9d52629c6a4c2ac"><span><div id="60617a5df4664afdb9d52629c6a4c2ac" class="notion-header-anchor"></div><a class="notion-hash-link" href="#60617a5df4664afdb9d52629c6a4c2ac" title="实验"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">实验</span></span></h4><div class="notion-text notion-block-5964332023434b479fc4f0c5790e6ad0">这篇文章4.2的实验很有意思，他们用T-SNE可视化退化表征模块（就是把它当做分类问题）</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-ea1ef870430647c6b709e3dafd2f1aac"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://raw.githubusercontent.com/The-Learning-And-Vision-Atelier-LAVA/DASR/main/Figs/fig.6.png?t=ea1ef870-4306-47c6-b709-e3dafd2f1aac" alt="notion image" loading="lazy" decoding="async"/></div></figure><div class="notion-text notion-block-c01536d77d84413aa22d6e21f1bee2e2">(a)是没用用退化表征模块的，(b)使用了。它表明该模块能够分辨出四种不同的退化核。</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-dc54b289951e40bf98c4bba996369617" data-id="dc54b289951e40bf98c4bba996369617"><span><div id="dc54b289951e40bf98c4bba996369617" class="notion-header-anchor"></div><a class="notion-hash-link" href="#dc54b289951e40bf98c4bba996369617" title="一些疑问"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">一些疑问</span></span></h4><div class="notion-text notion-block-3fc48964da78482d95ae75293e74e65b">这篇文章其实就是把对比学习想个法子直接套在SR问题上。这篇文章个人感觉存在以下问题：</div><ol start="1" class="notion-list notion-list-numbered notion-block-34b8718730eb44f3b8ee5cf29f042cc5"><li>一张图像上不同patch的退化都是一致的嘛？这个跟上一篇DASR的想法也有些出入，上一篇DASR认为非细节的信息在判别是否真实的过程中是近乎无用的，那么非细节的信息的退化方法可能就是简单的bicubic等，而细节方法就是很复杂的退化方案。</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-a72fba63b1094bb4a7b7b045bf559008"><li>文章使用卷积网络从LR得到退化的抽象表征，其实这种生成方案存在一定问题，我们假设LR = Func(SR)，那么我们怎么能够单从LR图像得到Func的信息呢？同时，这里可以GradCAM看看网络是如何得到退化抽象表示的。</li></ol><ol start="3" class="notion-list notion-list-numbered notion-block-2a656dd1bd68481c9fe901c38a3786d4"><li>如果遇到训练集没有的退化模型，这个退化表征是否仍然有效？</li></ol><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-8cfec091f3a245448d4b06feb4ad2a19" data-id="8cfec091f3a245448d4b06feb4ad2a19"><span><div id="8cfec091f3a245448d4b06feb4ad2a19" class="notion-header-anchor"></div><a class="notion-hash-link" href="#8cfec091f3a245448d4b06feb4ad2a19" title="CRL-SR"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">CRL-SR</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-c43bb672515d4bdfb5ca057af21e73c5" data-id="c43bb672515d4bdfb5ca057af21e73c5"><span><div id="c43bb672515d4bdfb5ca057af21e73c5" class="notion-header-anchor"></div><a class="notion-hash-link" href="#c43bb672515d4bdfb5ca057af21e73c5" title="文章简介"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">文章简介</span></span></h4><div class="notion-text notion-gray_background notion-block-7795df877c2044d789270f65e96a21f6">文章全称：Blind Image Super-Resolution via Contrastive Representation Learning</div><div class="notion-text notion-gray_background notion-block-7ca9cda73778479e8d705e4ccbe4e547">文章链接：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://ar5iv.labs.arxiv.org/html/2107.00708v1">https://ar5iv.labs.arxiv.org/html/2107.00708v1</a></div><div class="notion-text notion-gray_background notion-block-49cc9c0425b64940bd2aec373972f4c6">文章代码：null</div><div class="notion-text notion-block-b816708f8b674ff7b2acbc6e6c9b9c1c">这篇文章就是diss了上一篇文章的同一图像同一退化表征的问题，”We design CRL-SR, a contrastive representation learning network that focuses on blind SR of images <b>with multi-modal and spatially variant distributions</b>.”</div><div class="notion-text notion-block-648e0d4b30b6488cac389a1ea77a7ba9">文章的introduction部分说明了该文章的三个贡献：</div><ol start="1" class="notion-list notion-list-numbered notion-block-38170cb28ee84c21986fe48a53beb007"><li>offers a new and effective blind SR approach while images suffer from multi-modal, spatially variant, and unknown degradation.</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-07234a6cf5f746cfaa5ccd88d9779efc"><li>design a contrastive decoupling encoding technique that leverages feature contrast to extract resolution-invariant features across LR and HR images.</li></ol><ol start="3" class="notion-list notion-list-numbered notion-block-139bbc539a0f448aaac2e5180095007c"><li>design a conditional contrastive loss that guides to generate the high-frequency details that are lost during the image degradation process effectively.</li></ol><div class="notion-text notion-block-4b61e2c56276456cbb1ca0aa5242437b">实际来说，既然空域上退化表征存在变化，借鉴一下USR-DA中拉进学习到bicubic域与真实域LR的做法，我们拉进去掉低频&amp;去掉复杂退化的图像的距离就好了，这样就能够避免空域上不同退化对对比学习的影响</div><blockquote class="notion-quote notion-block-cdfde9f00156482db69cd054acd2fc4d"><div>为啥不利用不同patch的不同退化这一特点，反而要丢掉…</div></blockquote><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-7913d791a403408291b6b8e8dddc5d5a" data-id="7913d791a403408291b6b8e8dddc5d5a"><span><div id="7913d791a403408291b6b8e8dddc5d5a" class="notion-header-anchor"></div><a class="notion-hash-link" href="#7913d791a403408291b6b8e8dddc5d5a" title="解决方案"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">解决方案</span></span></h4><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-dd8708dd3d2d4c8681903eb62d89817d"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://ar5iv.labs.arxiv.org/html/2107.00708/assets/x1.png?t=dd8708dd-3d2d-4c86-8190-3eb62d89817d" alt="notion image" loading="lazy" decoding="async"/></div></figure><div class="notion-text notion-block-1c571308e2c84647b413abebc092d0ed">这一篇风格和思想都有点像USR-DA。</div><div class="notion-text notion-block-142038ceccc34518b9bbf52b942fc926">文章的核心解决思路：</div><div class="notion-text notion-block-f2f76df875b24fb998a6a64c77d866cc">CDE，上图的黄色部分：”which strives to disentangle and keep resolution-invariant features (i.e., clean low-frequency features)”</div><ul class="notion-list notion-list-disc notion-block-c7b0e50d9f8f4e5987feb5d3e8468e86"><li>El用于移除真实域LR图像的复杂退化，即把真实域LR图像特征映射到bicubic域LR图像特征</li></ul><ul class="notion-list notion-list-disc notion-block-1d1da20ad15848c3b48e8401cd1dc007"><li>Eh用于移除真实域LR图像的高频信息（或许可以用FFT代替）</li></ul><div class="notion-text notion-block-603acc53ecd34759a5f80b4a938ba49e">这里使用了NCELoss拉进fl与fh，这意味着正负样本的选取为在M个fl’组成的fl中存在一一对应的fh’，这里的fl’与fh’互为正样本，并与其他fl中的特征互为负样本</div><blockquote class="notion-quote notion-block-eec94de2ce254b1a80b6207b94e69b37"><div>这里需要注意，如果跟DASR一样选取不同图像的特征来做对比学习，会导致负样本过于简单，效果不好</div></blockquote><div class="notion-text notion-block-8da8f3f79cdc4be5830d00440cf5cc33">CFR，上图的绿色部分：”CFR is designed to recover high-frequency details that are lost or corrupted during the image degradation process.”</div><ul class="notion-list notion-list-disc notion-block-c8ce8c7d466e4a5da9084b0ede0552b2"><li>我们使用低频&amp;去除了复杂退化的fl输入超分网络中，得到的超分图像必然是缺失了很多高频信息的，需要在CFR过程中弥补回来</li></ul><ul class="notion-list notion-list-disc notion-block-deccfecb542d4ed895067eee95a35f9e"><li>用低频&amp;去除了复杂退化后超分的图像与真实的HR图像计算loss</li></ul><div class="notion-text notion-block-3fef87ef06c44146a40cf88a28f6dbad">这里仍然使用NCELoss，但是考虑到每张图像的高频信息不一致，作者在NCELoss的基础上增加了一个正则项</div><div class="notion-text notion-block-fc8ae54afb604bef9ce8af259677e41e">该正则项能够增强网络学到的fh与fl的差别，确保高频信息与bicubic域LR图像特征不一样（同时通过CDE，间接确保高频信息与去除了高频信息的图像的特征不一样）。</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-13440ce919ab47ed9d9edac70506fea0" data-id="13440ce919ab47ed9d9edac70506fea0"><span><div id="13440ce919ab47ed9d9edac70506fea0" class="notion-header-anchor"></div><a class="notion-hash-link" href="#13440ce919ab47ed9d9edac70506fea0" title="一些疑问"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">一些疑问</span></span></h4><ol start="1" class="notion-list notion-list-numbered notion-block-3a1d3533dd824250b073532c4eeb10a3"><li>高频信息的移除个人觉得用一个Encoder去学有点小题大做，或许可以试试直接JPEG等操作在输入一个Encoder。</li></ol><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-6cbd0dbf811148818246953428efdae7" data-id="6cbd0dbf811148818246953428efdae7"><span><div id="6cbd0dbf811148818246953428efdae7" class="notion-header-anchor"></div><a class="notion-hash-link" href="#6cbd0dbf811148818246953428efdae7" title="CDSR"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">CDSR</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-c98d598a3de14fa7979817e881627098" data-id="c98d598a3de14fa7979817e881627098"><span><div id="c98d598a3de14fa7979817e881627098" class="notion-header-anchor"></div><a class="notion-hash-link" href="#c98d598a3de14fa7979817e881627098" title="文章简介"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">文章简介</span></span></h4><div class="notion-text notion-gray_background notion-block-46ec18783cb04af6891b1fa1e4e0cde9">文章全称：Joint Learning Content and Degradation Aware Embedding for Blind Super-Resolution</div><div class="notion-text notion-gray_background notion-block-4dbcc1924ce041a3932c5cdae720a0b3">文章链接：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://ar5iv.labs.arxiv.org/html/2208.13436v1">https://ar5iv.labs.arxiv.org/html/2208.13436v1</a></div><div class="notion-text notion-gray_background notion-block-c7f19862c92d47d3ba9c8867956d227d">文章代码：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/ZhouYiiFeng/CDSR">https://github.com/ZhouYiiFeng/CDSR</a></div><blockquote class="notion-quote notion-block-458b486062704bf8a4cf2c853f96cdab"><div>这篇文章的introduction和motivation在它的github主页写的很清楚了，这里仅做一些翻译</div></blockquote><div class="notion-text notion-block-0268fa5ae5484fe0a037d8289e4c4295">目前的退化预测存在两种思路：</div><ol start="1" class="notion-list notion-list-numbered notion-block-dbc02c2310514df8846998e95a0cd490"><li>Supervised Kernel Prediction: 显式或者隐式的学习退化核，退化核的估计会更为准确但是无法适用所有场景（或者跟segment anything一样来个1.1B的退化核种类？）</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-4cdbfac91f7747069329aab11ab5db3b"><li>Unsupervised Degradation Prediction: 不需要显式或者隐式的估计退化，能够避免SKP只能处理特定模糊核的情况（DASR就是典型的代表）</li></ol><div class="notion-text notion-block-5cfb62a0acd848f49961185f7ee900da">SKP能够很好的解决训练集出现的所有退化，但是可能无法解决训练集以外的退化；UDP能够处理较多退化，但是可能难以招架复杂的退化。</div><div class="notion-text notion-block-8fdc890d523d490ba6e5b0f48e080c05">这篇文章作者为了弥补UDP的精度问题，发现了一个很奇怪的现象：”Surprisingly, we observed that using a degradation-oriented embedding will fail to improve the network performance or even end up with poor results.”，即使用面向退化的嵌入将无法改善网络性能。因为退化核和图像内容本身存在domain gap，强行将模糊核嵌入超分网络会损害性能，应当将内容与退化同时embedding。这里可能从IKC获得的启发，在IKC中提到退化空间和内容空间的融合会造成伪影（因为如果只将退化与图像叠加在一起，实际上这一步是在人为制造LR数据，虽然在PDM等文章里面给出了该问题的解法）。</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-763c71977230477597dcd5925cdde6d0" data-id="763c71977230477597dcd5925cdde6d0"><span><div id="763c71977230477597dcd5925cdde6d0" class="notion-header-anchor"></div><a class="notion-hash-link" href="#763c71977230477597dcd5925cdde6d0" title="分析"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">分析</span></span></h4><div class="notion-text notion-block-278ff15969fe4549b3846c51b9135690">文章专门分析了”<b><b>Degradation Embedding</b></b>”: “<b>The Higher Degradation Accuracy the Better?</b>”</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-ff4019854fc744098659388c64c05886"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:100%"><img style="object-fit:cover" src="https://raw.githubusercontent.com/hhhfccz/image/master/img/sft_o.png?t=ff401985-4fc7-4409-8659-388c64c05886" alt="notion image" loading="lazy" decoding="async"/></div></figure><div class="notion-text notion-block-cf4d3c311a484da0a600d30b47b51d21">上图左侧是DASR超分特征可视化，右边的特征显得更有层次感（可以看出内容信息被embedding进来了，而非只有退化带来的局部差异）。</div><div class="notion-text notion-block-ff733abb47c543e78a26d44a7b23b6ac">同时，本文提出的方案能够更使得学习到的退化表征有更丰富的内容信息，而非完全不一样（引入了内容信息之后，退化表征能够分布的更加分散而不展现出明显的曲线关系，这意味他们并不跟退化本身相关，还收到了其他因素影响，比如内容信息），如下图：</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-83809b6d417041c083519c711b68e306"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://raw.githubusercontent.com/hhhfccz/image/master/img/cdsr_x2.png?t=83809b6d-4170-41c0-8351-9c711b68e306" alt="notion image" loading="lazy" decoding="async"/></div></figure><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-54feeca1c2ab4cad9ded7af830db16d4" data-id="54feeca1c2ab4cad9ded7af830db16d4"><span><div id="54feeca1c2ab4cad9ded7af830db16d4" class="notion-header-anchor"></div><a class="notion-hash-link" href="#54feeca1c2ab4cad9ded7af830db16d4" title="解决方案"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">解决方案</span></span></h4><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-7139eda1fc234264897c188caddd09b8"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://ar5iv.labs.arxiv.org/html/2208.13436/assets/x4.png?t=7139eda1-fc23-4264-897c-188caddd09b8" alt="notion image" loading="lazy" decoding="async"/></div></figure><div class="notion-text notion-block-e0c8174f9f374154874e302e4442264d">文章的claim的贡献就是针对文章的解决方案：</div><ol start="1" class="notion-list notion-list-numbered notion-block-dc78acdd37454ba89ccd1d2f70c719d1"><li>a lightweight patch-based encoder (LPE) to extract content-aware degradation embedding features</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-b5e92df12cc746f4b66685f104410cb3"><li>DQA to adaptively fuse the predicted content and degradation aware embedding into the SR network</li></ol><ol start="3" class="notion-list notion-list-numbered notion-block-048ca9d244cb4fb99ed05c1aa76c799a"><li>a Codebook-based Space Compress module (CSC) to limit the basis of feature space.</li></ol><div class="notion-text notion-block-c27dc70f23254ee7af10d470958d3d95">其中LPE加入了Patch-Wise的语义信息用来增强内容在退化表征学习中的存在感，Pixel-Wise的退化特征提取抄的MANet。对比学习过程中仍然用的InfoNCE。</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-81409f8490844e1dbfac1f3493204d64" data-id="81409f8490844e1dbfac1f3493204d64"><span><div id="81409f8490844e1dbfac1f3493204d64" class="notion-header-anchor"></div><a class="notion-hash-link" href="#81409f8490844e1dbfac1f3493204d64" title="一些疑问"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">一些疑问</span></span></h4><ol start="1" class="notion-list notion-list-numbered notion-block-34130cf1f68549beafa65ee838efd05e"><li>在分析中，作者说明其方法能够关注更全局的信息，可以放进LAM中看看（不过LAM有个问题，它对退化十分敏感，不能解释真实场景下的超分模型，这里可能有个大坑）</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-b033bb9629ff4f3fbe0a40083c4a35df"><li>如果不是把内容和退化联合嵌入呢？比如：分割结果、检测结果、分类中的特征。</li></ol><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-740b7de1d8f945b4bc670a28603d2a12" data-id="740b7de1d8f945b4bc670a28603d2a12"><span><div id="740b7de1d8f945b4bc670a28603d2a12" class="notion-header-anchor"></div><a class="notion-hash-link" href="#740b7de1d8f945b4bc670a28603d2a12" title="DAA"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">DAA</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-aefa350f01f64079ad27ee9cbdd0a6ec" data-id="aefa350f01f64079ad27ee9cbdd0a6ec"><span><div id="aefa350f01f64079ad27ee9cbdd0a6ec" class="notion-header-anchor"></div><a class="notion-hash-link" href="#aefa350f01f64079ad27ee9cbdd0a6ec" title="文章简介"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">文章简介</span></span></h4><div class="notion-text notion-gray_background notion-block-544055457d1646259433503f2ebe7277">文章全称：Blind Image Super-Resolution with Degradation-Aware Adaptation</div><div class="notion-text notion-gray_background notion-block-cee72e8502d345759646c15a00cffe20">文章链接：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://openaccess.thecvf.com/content/ACCV2022/papers/Wang_Blind_Image_Super-Resolution_with_Degradation-Aware_Adaptation_ACCV_2022_paper.pdf">https://openaccess.thecvf.com/content/ACCV2022/papers/Wang_Blind_Image_Super-Resolution_with_Degradation-Aware_Adaptation_ACCV_2022_paper.pdf</a></div><div class="notion-text notion-gray_background notion-block-07e1fee17e2d474fa1f808ad1db60051">文章代码：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/wangyue7777/blindsr_daa">https://github.com/wangyue7777/blindsr_daa</a></div><div class="notion-text notion-block-d0871329149b400f9cfdf30bc1527729">这一篇文章提出了一个问题：在基于对比学习的DASR方案中，退化表征学习模块只能输出当前的退化表征，这会导致超分模块并不清楚当前输入的LR的退化复杂程度，而仅有真实LR与其对应的退化表征是不够的，还需要知道退化是否严重。</div><div class="notion-text notion-block-916830deadd144399612aab06c6ae2b2">这个motivation（或者说故事）比较难理解，我们用老师与学生做例子：对比学习中一次仅有一位老师（Decoder），两位学生（输入的LR图像），会有一名助教（Encoder）帮助老师（Decoder）为学生辅导（超分），助教（Encoder）会先比较两位学生（LR图像）的差别、对他们的情况（退化表征）进行总结，并将学生（LR图像）和自己的总结（退化表征）递给老师。这种情况下，老师（Decoder）一次只能看到两个学生（LR图像）的情况，并做出下一步判断（如何超分）。</div><div class="notion-text notion-block-a83caac0bbba4fdf845a16ad8a7ee73d">这种教学方法存在以下问题：</div><ol start="1" class="notion-list notion-list-numbered notion-block-9226106e333446b2b51532aacdf61769"><li>老师不知道学生的整体水平（所有数据集的情况）</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-3831ad3db69b4f709cc35c8d59a6bfab"><li>老师不知道单个学生在整体水平中的位次（本文的出发点）</li></ol><ol start="3" class="notion-list notion-list-numbered notion-block-56ed164f8fd94e9983aa4dfc05a7e7a4"><li>老师不知道单个学生的特性（这个问题在单纯的超分网络中不存在，但是在联合任务中，Encoder可能无法认知到噪声多还是模糊多，可能需要更复杂的Encoder设计）</li></ol><div class="notion-text notion-block-f177e11580dd4af4b7fbcab3323b2bcd">本文着重强调了第二点，同时引出（套用）了Ranking Loss作为解决方案。</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-1471a8512b56418c9b98d5e505db05ff" data-id="1471a8512b56418c9b98d5e505db05ff"><span><div id="1471a8512b56418c9b98d5e505db05ff" class="notion-header-anchor"></div><a class="notion-hash-link" href="#1471a8512b56418c9b98d5e505db05ff" title="解决方案"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">解决方案</span></span></h4><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-9c5cf73ff3de474f8f033d81747db05e"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column;height:100%"><img style="object-fit:cover" src="https://www.notion.so/image/https%3A%2F%2Fs3-us-west-2.amazonaws.com%2Fsecure.notion-static.com%2Fd47d5bf6-ff50-4187-93b1-2f3b9789d546%2Fdaa.png?table=block&amp;id=9c5cf73f-f3de-474f-8f03-3d81747db05e&amp;t=9c5cf73f-f3de-474f-8f03-3d81747db05e&amp;width=857.9921875&amp;cache=v2" alt="notion image" loading="lazy" decoding="async"/></div></figure><div class="notion-text notion-block-e3964a78378e438bb7b9c6a5563f6dc1">其中与DASR不一样的地方就在于Ranker这个模块，Ranker用于学习对退化表征进行排名（额外对Encoder增加了一个head）</div><div class="notion-text notion-block-c198b62536024c01b21987bc97b764d2">同时，本文也提出一个空域上自适应的退化表征嵌入超分网络的策略，本质上也是认为全图一致的退化是有问题的。</div><div class="notion-text notion-block-3e953798a7224e97975c3e5c0b6d87fc">即：</div><ol start="1" class="notion-list notion-list-numbered notion-block-4ea7763a3d454243bdaa4a84581778dc"><li>Ranking loss is imposed on top to make correct decision on estimating the degree of degradation</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-e101b4df88d745da9ad771a8737ec141"><li>The degradation is spatially-invariant, different types of textures may have different sensitivity to the degradation</li></ol><div class="notion-text notion-block-0a585323437e45b886246fa05ac68313">在RankSRGAN中，Ranker的作用是根据图像的感知得分对图像进行排名，而在本文Ranker的作用就是根据图像的Encoder结果对图像的退化模型难度进行排名，其他基本相同</div><blockquote class="notion-quote notion-block-bfc4b2feaa87407f82c68c4971b347c0"><div>同时，为了保证模型能够学习到不同退化图像的rank，数据的预处理也很重要，需要手动制作很多不同退化的数据，这一点是否对真实图像超分有利也有待商榷</div></blockquote><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-ef0e4a9bf5cd42419df26f50598665b3" data-id="ef0e4a9bf5cd42419df26f50598665b3"><span><div id="ef0e4a9bf5cd42419df26f50598665b3" class="notion-header-anchor"></div><a class="notion-hash-link" href="#ef0e4a9bf5cd42419df26f50598665b3" title="一些疑问"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">一些疑问</span></span></h4><div class="notion-text notion-block-bb589ccde825465f87e3f16007b9be7d">其实这篇文章的出发点还是比较新颖的，但是论文语言组织的不是很好、实验也不是很充足，缺少了很多超分SoTA模型的消融实验，我有以下疑问：</div><ol start="1" class="notion-list notion-list-numbered notion-block-e885d3aa5bc54e799173018feb17be43"><li>可能这篇文章仅在轻量化模型（或者说欠拟合的模型上）表现良好，作者仅对比了IMDN、RCAN、EDSR的消融实验，这是否说明这一方法在已经具有较强表现能力的超分网络上效果不佳</li></ol><h2 class="notion-h notion-h1 notion-h-indent-0 notion-block-233d223c024645d18836f7bfdd2eef82" data-id="233d223c024645d18836f7bfdd2eef82"><span><div id="233d223c024645d18836f7bfdd2eef82" class="notion-header-anchor"></div><a class="notion-hash-link" href="#233d223c024645d18836f7bfdd2eef82" title="LR Generator"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">LR Generator</span></span></h2><div class="notion-text notion-block-39914b8e8cd44255ad9d947bedd350dc">数据增强（得到了估计的模糊核或者退化表示后，将其作用于HR图像得到成对的训练数据），因为其可以做到非配对，所以也纳入无监督方案中。</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-299f450387c046da9843b344d42bcd99" data-id="299f450387c046da9843b344d42bcd99"><span><div id="299f450387c046da9843b344d42bcd99" class="notion-header-anchor"></div><a class="notion-hash-link" href="#299f450387c046da9843b344d42bcd99" title="DSGAN"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">DSGAN</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-5ad539a82af2443bb7c0a5b6a5fb12de" data-id="5ad539a82af2443bb7c0a5b6a5fb12de"><span><div id="5ad539a82af2443bb7c0a5b6a5fb12de" class="notion-header-anchor"></div><a class="notion-hash-link" href="#5ad539a82af2443bb7c0a5b6a5fb12de" title="文章简介"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">文章简介</span></span></h4><div class="notion-text notion-gray_background notion-block-3f7456d0586b4edbbbd30115f4e7586d">文章全称：Frequency Consistent Adaptation for Real World Super Resolution</div><div class="notion-text notion-gray_background notion-block-6c06b9e475444803bb531c92498bc0ee">文章链接：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://ar5iv.labs.arxiv.org/html/2012.10102v1">https://ar5iv.labs.arxiv.org/html/2012.10102v1</a></div><div class="notion-text notion-gray_background notion-block-6c83ed484df54fd68abf4aa18f98487c">文章代码：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/ManuelFritsche/real-world-sr">https://github.com/ManuelFritsche/real-world-sr</a></div><blockquote class="notion-quote notion-block-427646a4801045fc9b29c2cab48955c8"><div><a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://ieeexplore.ieee.org/abstract/document/9150978/">Guided Frequency Separation Network for Real-World Super-Resolution</a>也是一篇类似的工作，基本思想类似</div></blockquote><div class="notion-text notion-block-e33e919befff494e90d90fb9a07b5ff8">这篇文章的出发点为：”The domain gap between the LR images and the real-world images can be observed clearly on<b> frequency density</b>, which inspires us to explictly narrow the undesired gap caused by incorrect degradation.”作者发现在不同退化程度的图像之间的差异在频率域会更加明显，如下图：</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-0cb4856862bf4e88a24868ce5d31f986"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://ar5iv.labs.arxiv.org/html/2012.10102/assets/x4.png?t=0cb48568-62bf-4e88-a248-68ce5d31f986" alt="notion image" loading="lazy" decoding="async"/></div></figure><div class="notion-text notion-block-0aab11c94fc440b18ce6f97646b09972">这张图含义如下：</div><ol start="1" class="notion-list notion-list-numbered notion-block-21a4a8d483fa4ac89d99185df7bec904"><li>(a)为不同退化模糊核下图像的频率密度，模糊核本身方差越大（A &gt; B ≈ Source &gt; C），得到的图像的频率密度越小。</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-d9904b4664ed48bdb85d00c18d110980"><li>(b)为上下采样的区别，对图像进行下采样后，图像的频率密度将变高；反之，上采样后，图像的频率密度将变小。</li></ol><div class="notion-text notion-block-fe462ba962424109a648528b0991e706">基于此规律，作者认为”The relationship between degradation and frequency density motivates us to keep frequency consistency between 𝐈_LR and source image 𝐱 . We focus on estimating 𝐤 with frequency domain regularization”模糊核的预测，应当受到频率的指导。本文提出的指导是一种领域迁移（image transfer）的想法。</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-b2d92dd443b443f0bb2427086a1231fe" data-id="b2d92dd443b443f0bb2427086a1231fe"><span><div id="b2d92dd443b443f0bb2427086a1231fe" class="notion-header-anchor"></div><a class="notion-hash-link" href="#b2d92dd443b443f0bb2427086a1231fe" title="解决方案"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">解决方案</span></span></h4><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-bd82378df07045018ec05921ba7289b1"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://ar5iv.labs.arxiv.org/html/2012.10102/assets/x5.png?t=bd82378d-f070-4501-8ec0-5921ba7289b1" alt="notion image" loading="lazy" decoding="async"/></div></figure><div class="notion-text notion-block-6d0a555ec5f04db080a31ab2bb76cd3c">本文的LR-HR对中，LR的图像制作是从bicubic图像迁移的，迁移过程纳入了前文提到的频率密度先验。本质上是半个CycleGAN，判别器使用的与第一个DASR一样的小波判别器。</div><div class="notion-text notion-block-b66a7f03cf034f2689e7bc0ac36dcadd">同期存在一篇<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://ar5iv.labs.arxiv.org/html/2110.12151v1">Spectrum-to-Kernel Translation for Accurate Blind Image Super-Resolution</a>文章也从证明了从频率域学习退化的有效性。遗憾的是，这两篇文章都只是从实验上证明了这一点。</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-89e7e6cacece4660837ffd1cec3e0d69" data-id="89e7e6cacece4660837ffd1cec3e0d69"><span><div id="89e7e6cacece4660837ffd1cec3e0d69" class="notion-header-anchor"></div><a class="notion-hash-link" href="#89e7e6cacece4660837ffd1cec3e0d69" title="一些疑问"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">一些疑问</span></span></h4><ol start="1" class="notion-list notion-list-numbered notion-block-27d7f536065b4ac2ae3348205a71f6df"><li>从(a)中看，方差越小，模糊核之间的频率密度差异越小（越难以分辨），或许可以结合标准差等统计参数一起比较</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-ec00a837f71f4ee4bcbeb8db953e78c6"><li>老问题，判别器不应该只输入高频分量（在<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://ieeexplore.ieee.org/abstract/document/9150978">Guided Frequency Separation Network for Real-World Super-Resolution</a>一文中有解决方案，就是一个减法）</li></ol><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-93e25b2ed3f34bd995449ce58f2236be" data-id="93e25b2ed3f34bd995449ce58f2236be"><span><div id="93e25b2ed3f34bd995449ce58f2236be" class="notion-header-anchor"></div><a class="notion-hash-link" href="#93e25b2ed3f34bd995449ce58f2236be" title="MSSR"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">MSSR</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-110e10315ff4481d8990b6e4712fe607" data-id="110e10315ff4481d8990b6e4712fe607"><span><div id="110e10315ff4481d8990b6e4712fe607" class="notion-header-anchor"></div><a class="notion-hash-link" href="#110e10315ff4481d8990b6e4712fe607" title="文章简介"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">文章简介</span></span></h4><div class="notion-text notion-gray_background notion-block-2e5b329046394015b4a9381d79b1d677">文章全称：Learning the Degradation Distribution for Blind Image Super-Resolution</div><div class="notion-text notion-gray_background notion-block-1ac021b392c942b59c5c8303778b3bde">文章链接：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://ar5iv.labs.arxiv.org/html/2201.10747v2">https://ar5iv.labs.arxiv.org/html/2201.10747v2</a></div><div class="notion-text notion-gray_background notion-block-60f498a9ce5c4008809f19ce87727fe2">文章代码：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/sangyun884/MSSR">https://github.com/sangyun884/MSSR</a></div><div class="notion-text notion-block-1c333d74b7a64119a57be7ef193e945c">这篇文章专注于解决训练一个更真实的退化图像生成器，作者认为：”previous works have heavily relied on the unrealistic assumption that the conditional distribution is a delta function and learned the deterministic mapping from the HR image to a LR image”。也就是以前的工作，只能从一个bicubic图像引导生成出一个确定的真实LR图像，这不是很病态（超分问题应该存在病态问题，而不是一一对应关系）。</div><div class="notion-text notion-block-4e1974274e564ec18e41a4d3b0d2a963">为此，文章提出了一个十分简单的解决方案：”add a Gaussian noise multiplied by the learned standard deviation to intermediate feature maps of the degradation generator”。直接在特征层面引入噪声。</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-e9458e239abf4d6bb5fd93bc91bd7175" data-id="e9458e239abf4d6bb5fd93bc91bd7175"><span><div id="e9458e239abf4d6bb5fd93bc91bd7175" class="notion-header-anchor"></div><a class="notion-hash-link" href="#e9458e239abf4d6bb5fd93bc91bd7175" title="解决方案"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">解决方案</span></span></h4><div class="notion-text notion-block-777fab0b017946f89fcdcbda8e10f941">文章的解决方案确实比较simple but powerful。</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-b59a09cd8e824ca89050e5ad3c72ac24"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://ar5iv.labs.arxiv.org/html/2201.10747/assets/figure/overview.jpg?t=b59a09cd-8e82-4ca8-9050-e5ad3c72ac24" alt="notion image" loading="lazy" decoding="async"/></div></figure><div class="notion-text notion-block-682678679482416e9d5ffe74f5588e36">G1到Gk就是k个下采样网络，M1到Mk就是k个噪声添加器，同时”it is crucial for each generator to exclusively cover a certain area of the p(𝐱|𝐲) without redundancy”多个退化生成器可以带来更丰富的LR信息，使得网络能够处理更广泛的退化。最终构成的LR-HR对中，LR为k个退化后图像的平均值。<div class="notion-text-children"><div class="notion-text notion-block-c603c400f8de4c799c6354846def3d98">同时为了避免”Given the limited capacity of a model, it can be problematic as we train a model to invert the multiple degradation generators.”作者使用了Collaborative Learning（协作学习），也就是图中的Lcol的由来，它促使不同的LR图像互相不一致，这样避免了不同的G与M组合后会学到一致的退化（作者在这里提到”We initially tried to apply knowledge distillation in feature space and found that both yield similar results.”图像层面的蒸馏和特征层面的蒸馏效果差不多）</div></div></div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-c5631690d8464f5f8f5682978e30c735" data-id="c5631690d8464f5f8f5682978e30c735"><span><div id="c5631690d8464f5f8f5682978e30c735" class="notion-header-anchor"></div><a class="notion-hash-link" href="#c5631690d8464f5f8f5682978e30c735" title="Noise Injection between StyleGAN and this paper"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">N<b><b>oise Injection between StyleGAN and this paper</b></b></span></span></h4><div class="notion-text notion-block-5150342f09b048b6a56109e996910488">可以看到，这个公式右侧第二项跟StyleGAN的噪声是一样的。</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-e9fdfe92549c4e9d819e0653e644fb10" data-id="e9fdfe92549c4e9d819e0653e644fb10"><span><div id="e9fdfe92549c4e9d819e0653e644fb10" class="notion-header-anchor"></div><a class="notion-hash-link" href="#e9fdfe92549c4e9d819e0653e644fb10" title="一些疑问"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">一些疑问</span></span></h4><ol start="1" class="notion-list notion-list-numbered notion-block-5e6bc45c53aa4ff4ba3693862f970f53"><li>在特征上直接引入噪声是否会难以训练，类似分类中mixup和cutmix的几个变体，原始的直接在图像操作，后续出来几个在特征层面操作的，或许可以借鉴一下。</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-c7b9d4dfcaa54babad798be6488d21be"><li>noise injection in GANs 是个大坑，这篇文章在超分问题中开了个头。</li></ol><ol start="3" class="notion-list notion-list-numbered notion-block-fb8aee96f527493cbe2bcf67d83d9f25"><li>多个退化生成器是不是不够efficient，一个退化生成器输出很多不同的退化，这样可以避免引入额外的Collaborative Learning</li></ol><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-7fe94480b0e1449dad0ec7a6ad06f374" data-id="7fe94480b0e1449dad0ec7a6ad06f374"><span><div id="7fe94480b0e1449dad0ec7a6ad06f374" class="notion-header-anchor"></div><a class="notion-hash-link" href="#7fe94480b0e1449dad0ec7a6ad06f374" title="PDM"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">PDM</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-bbdf6fbc38d7434e81f576a5e860b239" data-id="bbdf6fbc38d7434e81f576a5e860b239"><span><div id="bbdf6fbc38d7434e81f576a5e860b239" class="notion-header-anchor"></div><a class="notion-hash-link" href="#bbdf6fbc38d7434e81f576a5e860b239" title="文章简介"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">文章简介</span></span></h4><div class="notion-text notion-gray_background notion-block-9d262f7bd6bf4af48c5ff0c8ee6a59de">文章全称：Learning the Degradation Distribution for Blind Image Super-Resolution</div><div class="notion-text notion-gray_background notion-block-97873918f70a4fc29c8077fa2eeacbb3">文章链接：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://ar5iv.labs.arxiv.org/html/2203.04962v1">https://ar5iv.labs.arxiv.org/html/2203.04962v1</a></div><div class="notion-text notion-gray_background notion-block-5946ed389d684fe087d47d16daee2313">文章代码：<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/greatlog/UnpairedSR">https://github.com/greatlog/UnpairedSR</a></div><div class="notion-text notion-block-828909508e0c44d98e23de78d72593d3">这一篇更是纯粹：退化核跟LR图像无关，所有的模糊核估计就是一种更复杂的数据增强方案。</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-9e82b942827b4a2db9819ef23ec6645d" data-id="9e82b942827b4a2db9819ef23ec6645d"><span><div id="9e82b942827b4a2db9819ef23ec6645d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#9e82b942827b4a2db9819ef23ec6645d" title="解决方案"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">解决方案</span></span></h4><div class="notion-text notion-block-074b108dccc344dab28117aabcb2c815">这篇文章沿用了先下采样</div><figure class="notion-asset-wrapper notion-asset-wrapper-image notion-block-e9aba3092dbd4c7ab74d76aba360525b"><div style="position:relative;display:flex;justify-content:center;align-self:center;width:100%;max-width:100%;flex-direction:column"><img style="object-fit:cover" src="https://ar5iv.labs.arxiv.org/html/2203.04962/assets/x2.png?t=e9aba309-2dbd-4c7a-b74d-76aba360525b" alt="notion image" loading="lazy" decoding="async"/></div></figure><ol start="1" class="notion-list notion-list-numbered notion-block-48a730f23bae424897711f082acd01d5"><li>kernel从一个噪声信号中直接学习得到（softmax，确保和为1）</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-4945cfbe7b5041cf8058f21ed2ad3432"><li>noise从LR后的图像与噪声信号中学习得到（为了确保0均值假设，会有个正则损失）</li></ol><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-c2041d36444d4924893c40e6dff74b92" data-id="c2041d36444d4924893c40e6dff74b92"><span><div id="c2041d36444d4924893c40e6dff74b92" class="notion-header-anchor"></div><a class="notion-hash-link" href="#c2041d36444d4924893c40e6dff74b92" title="一些疑问"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">一些疑问</span></span></h4><blockquote class="notion-quote notion-block-c56a241f75a44493829873bf5ad82800"><div>说实话这篇文章还能在之前卷翻天的盲超分上提点明显，有点意料之外</div></blockquote><ol start="1" class="notion-list notion-list-numbered notion-block-f1d112bf44cc452f954dde8c066b4c16"><li>该模糊核生成方案没有办法确保模糊核的稀疏性（需要一些正则规范一下）</li></ol><ol start="2" class="notion-list notion-list-numbered notion-block-84dfc782ed80483e9fdf4859f1b2384a"><li>跟以前从噪声学GAN的问题一样，每次都得重新学一下退化</li></ol></main></div>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[OneFlow’s FLOPs]]></title>
            <link>https://blog.jongkhu.com//article/oneflow-flops</link>
            <guid>https://blog.jongkhu.com//article/oneflow-flops</guid>
            <pubDate>Thu, 19 Jan 2023 00:00:00 GMT</pubDate>
            <description><![CDATA[计算OneFlow深度学习框架模型的FLOPs]]></description>
            <content:encoded><![CDATA[<div id="notion-article" class="mx-auto overflow-hidden "><main class="notion light-mode notion-page notion-block-00d0a79ccf824cdf823aa9f739e30827"><div class="notion-viewport"></div><div class="notion-collection-page-properties"></div><h2 class="notion-h notion-h1 notion-h-indent-0 notion-block-78865b603b4b447d8755175d33bd1f97" data-id="78865b603b4b447d8755175d33bd1f97"><span><div id="78865b603b4b447d8755175d33bd1f97" class="notion-header-anchor"></div><a class="notion-hash-link" href="#78865b603b4b447d8755175d33bd1f97" title="flowflops: OneFlow 模型的 Flops 计算"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">flowflops: OneFlow 模型的 Flops 计算</span></span></h2><div class="notion-text notion-block-84ad8a398d81441fa40d26d2907b606d">用于计算 OneFlow 模型的 FLOPs 和 Parameters 的第三方库。</div><div class="notion-text notion-block-118b8b4f17764988bbb30d5c6bce00a4">源码地址: <code class="notion-inline-code">https://github.com/Oneflow-Inc/flow-OpCounter</code></div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-87deb0442aa247f58c4a5820c57b6d40" data-id="87deb0442aa247f58c4a5820c57b6d40"><span><div id="87deb0442aa247f58c4a5820c57b6d40" class="notion-header-anchor"></div><a class="notion-hash-link" href="#87deb0442aa247f58c4a5820c57b6d40" title="介绍 &amp; 使用"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">介绍 &amp; 使用</span></span></h3><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-cf983a7265c04774b28ba37205bd5e53" data-id="cf983a7265c04774b28ba37205bd5e53"><span><div id="cf983a7265c04774b28ba37205bd5e53" class="notion-header-anchor"></div><a class="notion-hash-link" href="#cf983a7265c04774b28ba37205bd5e53" title="FLOPs"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">FLOPs</span></span></h4><div class="notion-text notion-block-758042322a5f4ff2bc524c808b417c1e">有许多人分不清楚 FLOPs 和 MACs 之间的关系，如<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/sovrasov/flops-counter.pytorch/issues/70">ptflops中的issue</a></div><div class="notion-text notion-block-63492bf950b44528819f56317ccab5cc">针对该问题，可以查看<a target="_blank" rel="noopener noreferrer" class="notion-link" href="https://github.com/Lyken17/pytorch-OpCounter/blob/master/benchmark/README.md">thop中的解释</a>，翻译如下：</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-7153efe0d13b4e7cb7b46f6f9680eaff" data-id="7153efe0d13b4e7cb7b46f6f9680eaff"><span><div id="7153efe0d13b4e7cb7b46f6f9680eaff" class="notion-header-anchor"></div><a class="notion-hash-link" href="#7153efe0d13b4e7cb7b46f6f9680eaff" title="MACs, FLOPs, what is the difference?"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">MACs, FLOPs, what is the difference?</span></span></h4><div class="notion-text notion-block-57c154a64d3d465789227c354bbdd680"><code class="notion-inline-code">FLOPs</code> 是<b>浮动算子</b>(floating operations)的缩写，包括mul / add / div …等。</div><div class="notion-text notion-block-d6738dfa73d648f5b2df76846507e612"><code class="notion-inline-code">MACs</code> 代表执行的<b>乘法累加运算</b>，例如: <code class="notion-inline-code">a &lt;- a + (b x c)</code>。</div><div class="notion-text notion-block-13a78d95376b4a7fa1537e8ec9bbdc53">如文中所示，一<code class="notion-inline-code">MACs</code>有一<code class="notion-inline-code">mul</code>和一<code class="notion-inline-code">add</code>。这就是为什么在许多地方<code class="notion-inline-code">FLOPs</code>几乎是两倍<code class="notion-inline-code">MACs</code>的原因。</div><div class="notion-text notion-block-67cec381c463452ea3f0cf0afa3c48ca">然而，现实世界中的应用要复杂得多。让我们考虑一个矩阵乘法示例。<code class="notion-inline-code">A</code>是一个形状为 <em>m</em> × <em>n</em> 的矩阵，<code class="notion-inline-code">B</code>是一个 <em>n</em> × 1 的向量。</div><div class="notion-text notion-block-19d517912aab4a608b304c20f14b7cf7">它会是m*n个<code class="notion-inline-code">MACs</code>和2m*n个<code class="notion-inline-code">FLOPs</code>。但是这样的矩阵乘法实现速度很慢，需要并行化才能运行得更快。</div><div class="notion-text notion-block-91cdaf0401d647d58b96f12ead10d5e9">此时<code class="notion-inline-code">MACs</code>数值不再是 m*n 。</div><div class="notion-text notion-block-97468be6995c4ce9b10b906e9b915eba">在比较 MAC / FLOP 时，我们希望数字与实现无关并且尽可能通用。因此在 (thop)[https://github.com/Lyken17/pytorch-OpCounter] 中，<b>我们只考虑乘法的次数</b>，而忽略所有其他操作。</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-06ee405635c741fc83cb9e4dc2501bb8" data-id="06ee405635c741fc83cb9e4dc2501bb8"><span><div id="06ee405635c741fc83cb9e4dc2501bb8" class="notion-header-anchor"></div><a class="notion-hash-link" href="#06ee405635c741fc83cb9e4dc2501bb8" title="安装方法"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">安装方法</span></span></h4><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-4d6cfca091dc41fdbe076286ffce5e1d" data-id="4d6cfca091dc41fdbe076286ffce5e1d"><span><div id="4d6cfca091dc41fdbe076286ffce5e1d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#4d6cfca091dc41fdbe076286ffce5e1d" title="使用方法"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">使用方法</span></span></h4><div class="notion-text notion-block-4b25c7c5fc224b72867967cd5a397346">目前支持两种 FLOPs 计算策略： 在 Eager 模式下计算和在 Graph 模式下计算。</div><blockquote class="notion-quote notion-block-eb56c0f4c9d24f88a2b912a3fb67586e"><div>在 Graph 模式下计算耗时较长，但结果更加精确</div></blockquote><div class="notion-text notion-block-adaaa98e3eae4cdf95902cefb123300d">示例:</div><div class="notion-text notion-block-174ecbd190a64c63b1f82465b94febeb">输出:</div><div class="notion-text notion-block-948195828a794c378e79906ccec2685b">可以看到两种计算方式下的输出有一定差别，这是因为在 ResNet 的 forward 代码里存在类似 <code class="notion-inline-code">out += identity</code> 的语句，这会造成 FLOPs 额外增加。而在 Eager 模式下我们只关注在 <code class="notion-inline-code">__init__()</code> 中定义的网络层，所以这种情况不会在 Eager 模式中被 hook 到。</div><div class="notion-text notion-block-e1729a66d01748bd91e7007e78a1d748">我们可以计算一下有哪些 add_n 算子在 Eager 模式中被我们忽略了:</div><div class="notion-text notion-block-8ef2f02f4557453b82de2bb8806e095e">一共为 5,519,360 ，刚好为两种模式的输出差值 <code class="notion-inline-code">4127444456 - 4121925096 = 5519360</code></div><div class="notion-text notion-block-adf36336c5e94f11b518ee614caab55e">在 Eager 模式下也会存在一些小误差，一般认为 ResNet50 的 FLOPs 为 4.09G ，而这里计算得到 4.12G ，是因为一般研究中会忽略类似 ReLU 等算子的 FLOPs 计算，所以与真实数值会有一定误差。有关一般都忽略了哪些算子的计算，可以查看 <code class="notion-inline-code">fvcore</code> 的输出。</div><blockquote class="notion-quote notion-block-fffdcd56a5174b5d86dbe48f20589a27"><div>在ptflops包中也存在这样的问题，笔者也有在issue中回复，详见issue: https://github.com/sovrasov/flops-counter.pytorch/issues/94</div></blockquote><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-a9b3a57648ab48fc9cf118b0d798247f" data-id="a9b3a57648ab48fc9cf118b0d798247f"><span><div id="a9b3a57648ab48fc9cf118b0d798247f" class="notion-header-anchor"></div><a class="notion-hash-link" href="#a9b3a57648ab48fc9cf118b0d798247f" title="Eager &amp; Graph 模式下的 Flops 计算"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Eager &amp; Graph 模式下的 Flops 计算</span></span></h3><div class="notion-text notion-block-c6fcab9ae8f249f993a188d5aa9e33df">接下来我们以简单修改后的 ResNet18 中的 BasicBlock 为例介绍一下两种 FLOPs 计算方式，设定网络如下：</div><blockquote class="notion-quote notion-block-74fca3c3083f40fd99435f08c018dfa6"><div>我们统一假定输入形状为(1, 32, 64, 64)</div></blockquote><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-432a3874bd074bdea05dc93de22bd431" data-id="432a3874bd074bdea05dc93de22bd431"><span><div id="432a3874bd074bdea05dc93de22bd431" class="notion-header-anchor"></div><a class="notion-hash-link" href="#432a3874bd074bdea05dc93de22bd431" title="Eager"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Eager</span></span></h4><div class="notion-text notion-block-183eecbaf0934011ab5594b7b9c182cc">在 Eager 模式中，我们只关注 <code class="notion-inline-code">__init__()</code> 中定义的网络层，也就是</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-9c1a260a84914001a5f980923abad29e" data-id="9c1a260a84914001a5f980923abad29e"><span><div id="9c1a260a84914001a5f980923abad29e" class="notion-header-anchor"></div><a class="notion-hash-link" href="#9c1a260a84914001a5f980923abad29e" title="二维卷积"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">二维卷积</span></span></h4><div class="notion-text notion-block-22536591a6474ba9aeeabda58c5158f6">卷积的原理在此不再赘述，直接给出计算公式: 2<em>k</em>2 × <em>Hout</em> × <em>Wout</em> × <em>Cin</em> × <em>Cout</em> ÷ <em>Groups</em></div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-02419c5c427043d0b39171520009d8bf" data-id="02419c5c427043d0b39171520009d8bf"><span><div id="02419c5c427043d0b39171520009d8bf" class="notion-header-anchor"></div><a class="notion-hash-link" href="#02419c5c427043d0b39171520009d8bf" title="归一化"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">归一化</span></span></h4><div class="notion-text notion-block-84e324c9c5d9445ea29167f3682dce2a"><code class="notion-inline-code">batchnorm</code> 主要计算了均值、方差，并基于此对特征进行归一化与仿射变换，其 FLOPs 为 2 × <em>C</em> × <em>H</em> × <em>W</em></div><div class="notion-text notion-block-9d57ceccbcb443518fa7c426445f71d3">如果不进行仿射变换，则其 FLOPs 为 <em>C</em> × <em>H</em> × <em>W</em></div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-809d9403496e45a78aeeb3628e40e9f4" data-id="809d9403496e45a78aeeb3628e40e9f4"><span><div id="809d9403496e45a78aeeb3628e40e9f4" class="notion-header-anchor"></div><a class="notion-hash-link" href="#809d9403496e45a78aeeb3628e40e9f4" title="激活函数"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">激活函数</span></span></h4><div class="notion-text notion-block-505e593412984cb8b3ecb1856b2ee952"><code class="notion-inline-code">relu</code> 对输入(1, C, H, W)进行了 <code class="notion-inline-code">y = x if x &gt; 0 else 0</code> 操作，也就是其 FLOPs 为 <em>C</em> × <em>H</em> × <em>W</em></div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-0aff6825f3ea4260934f9539d4fbcbb7" data-id="0aff6825f3ea4260934f9539d4fbcbb7"><span><div id="0aff6825f3ea4260934f9539d4fbcbb7" class="notion-header-anchor"></div><a class="notion-hash-link" href="#0aff6825f3ea4260934f9539d4fbcbb7" title="线性层"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">线性层</span></span></h4><div class="notion-text notion-block-fd374af544f94b879fdc44db739021d0">线性层输入为 (1, C, H, W)</div><div class="notion-text notion-block-8ea3cf9d52a345ec83ae43109d9fc5f3">线性层权重为 (W1, W)</div><div class="notion-text notion-block-dc0a7cf62e524cf0ab5c79edc26b7dde">两者相乘的 FLOPs 为 <em>C</em> × <em>H</em> × <em>W</em>1 × <em>W</em></div><div class="notion-text notion-block-44def485334b42048e8c1e326b48c790">其本质与 <code class="notion-inline-code">matmul</code> 计算相当</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-dd72723f4f424972b0e017620ee8dc03" data-id="dd72723f4f424972b0e017620ee8dc03"><span><div id="dd72723f4f424972b0e017620ee8dc03" class="notion-header-anchor"></div><a class="notion-hash-link" href="#dd72723f4f424972b0e017620ee8dc03" title="Graph"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Graph</span></span></h4><div class="notion-text notion-block-fcd790b3c17c41b79a2be46eea47a18f">在 Graph 模式中，我们会将 <code class="notion-inline-code">flow.nn.Module</code> 编译为 <code class="notion-inline-code">flow.nn.Graph</code> ，从 <code class="notion-inline-code">Graph</code> 中抽取出每一个算子输入的张量形状后再对网络的 FLOPs 进行计算</div><div class="notion-text notion-block-dea574f5e27b4880a8d67760912a899b">上述网络转换后的 Graph:</div><div class="notion-text notion-block-1594c6d35681492092ef0b3a3a631044"><code class="notion-inline-code">Graph</code> 中由 <code class="notion-inline-code">OPERATOR</code> 开始的层就是我们所需要的信息</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-19910d57f32241bba539890d94d0137d" data-id="19910d57f32241bba539890d94d0137d"><span><div id="19910d57f32241bba539890d94d0137d" class="notion-header-anchor"></div><a class="notion-hash-link" href="#19910d57f32241bba539890d94d0137d" title="卷积"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">卷积</span></span></h4><div class="notion-text notion-block-ded21d003fed4382b397cfbe919d8d94">在 <code class="notion-inline-code">flow.nn.Graph</code> 中 <code class="notion-inline-code">conv3x3</code> 和 <code class="notion-inline-code">conv1x1</code> 会被拆解为 <code class="notion-inline-code">conv2d + bias_add(if bias==True)</code></div><div class="notion-text notion-block-c9689604d20d4706a9fdc3ea30ef180b">由于我们只关注的卷积层的输入，而在计算 FLOPs 时需要得到卷积层输出的特征尺寸，所以我们需要依据输入计算输出特征的分辨率，方法如下</div><div class="notion-text notion-block-4a70618739f04222ae94abfe585401e0">随后即可正常计算 FLOPs</div><blockquote class="notion-quote notion-block-e884482d8f2442c3928576d34235e9b4"><div>至于为什么不直接得到算子输出的形状，因为解析 Graph 需要占用更多的额外时间</div></blockquote><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-de8710ff32034904b88dfe82d5ff4fcd" data-id="de8710ff32034904b88dfe82d5ff4fcd"><span><div id="de8710ff32034904b88dfe82d5ff4fcd" class="notion-header-anchor"></div><a class="notion-hash-link" href="#de8710ff32034904b88dfe82d5ff4fcd" title="归一化"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">归一化</span></span></h4><div class="notion-text notion-block-b6a229ff9b924555801df34f0e3d7e0d">在 <code class="notion-inline-code">flow.nn.Graph</code> 中 <code class="notion-inline-code">norm_layer(bn)</code> 是一个单独的算子，其计算方法与 Eager 模式中保持一致</div><blockquote class="notion-quote notion-block-45ee1dd3130f4dd594b0db4a3b5c0403"><div>需要注意的是 InstanceNorm 和 GroupNorm 在 flow.nn.Graph 中将被拆解为若干胶水算子，需要逐个计算</div></blockquote><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-f9e6747b8e234a9d9f0c63aae3223177" data-id="f9e6747b8e234a9d9f0c63aae3223177"><span><div id="f9e6747b8e234a9d9f0c63aae3223177" class="notion-header-anchor"></div><a class="notion-hash-link" href="#f9e6747b8e234a9d9f0c63aae3223177" title="激活函数"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">激活函数</span></span></h4><div class="notion-text notion-block-e0670b7c73244c84871316304f364c42">在 <code class="notion-inline-code">flow.nn.Graph</code> 中 <code class="notion-inline-code">relu</code> 是一个单独的算子，其 FLOPs 计算方法与 Eager 模式中保持一致</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-5d092dfa626e423c89c060495c826cae" data-id="5d092dfa626e423c89c060495c826cae"><span><div id="5d092dfa626e423c89c060495c826cae" class="notion-header-anchor"></div><a class="notion-hash-link" href="#5d092dfa626e423c89c060495c826cae" title="线性层"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">线性层</span></span></h4><div class="notion-text notion-block-39a9c7c88e854af6a6b2f245cbe1f8b8">在 <code class="notion-inline-code">flow.nn.Graph</code> 中 <code class="notion-inline-code">linear</code> 会被拆解为 <code class="notion-inline-code">matmul + broadcast_add(if bias==True)</code>，其 FLOPs 计算公式与 Eager 模式中基本一致</div><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-599628df46b34ec2b3285f747411c5a2" data-id="599628df46b34ec2b3285f747411c5a2"><span><div id="599628df46b34ec2b3285f747411c5a2" class="notion-header-anchor"></div><a class="notion-hash-link" href="#599628df46b34ec2b3285f747411c5a2" title="目前支持的 Op 与模型"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">目前支持的 Op 与模型</span></span></h3><div class="notion-text notion-block-3f6df0c6dca0441e821a9ec5acdcbaa3">目前该工具支持绝大部分算子、网络层与大多数 CNN ，列表如下</div><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-ab7e3239b21240dfb1e475250ffa0bcd" data-id="ab7e3239b21240dfb1e475250ffa0bcd"><span><div id="ab7e3239b21240dfb1e475250ffa0bcd" class="notion-header-anchor"></div><a class="notion-hash-link" href="#ab7e3239b21240dfb1e475250ffa0bcd" title="Eager"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Eager</span></span></h4><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-c733dd98bb5748b0ba4752877b381381" data-id="c733dd98bb5748b0ba4752877b381381"><span><div id="c733dd98bb5748b0ba4752877b381381" class="notion-header-anchor"></div><a class="notion-hash-link" href="#c733dd98bb5748b0ba4752877b381381" title="Graph"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">Graph</span></span></h4><h4 class="notion-h notion-h3 notion-h-indent-2 notion-block-09eb5fb9c4e1422186cba2c5e3f57d7b" data-id="09eb5fb9c4e1422186cba2c5e3f57d7b"><span><div id="09eb5fb9c4e1422186cba2c5e3f57d7b" class="notion-header-anchor"></div><a class="notion-hash-link" href="#09eb5fb9c4e1422186cba2c5e3f57d7b" title="FlowVision 中部分模型的计算结果"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">FlowVision 中部分模型的计算结果</span></span></h4><h3 class="notion-h notion-h2 notion-h-indent-1 notion-block-da94bf8cf6a44cbb9d2cae8675e8524c" data-id="da94bf8cf6a44cbb9d2cae8675e8524c"><span><div id="da94bf8cf6a44cbb9d2cae8675e8524c" class="notion-header-anchor"></div><a class="notion-hash-link" href="#da94bf8cf6a44cbb9d2cae8675e8524c" title="总结"><svg viewBox="0 0 16 16" width="16" height="16"><path fill-rule="evenodd" d="M7.775 3.275a.75.75 0 001.06 1.06l1.25-1.25a2 2 0 112.83 2.83l-2.5 2.5a2 2 0 01-2.83 0 .75.75 0 00-1.06 1.06 3.5 3.5 0 004.95 0l2.5-2.5a3.5 3.5 0 00-4.95-4.95l-1.25 1.25zm-4.69 9.64a2 2 0 010-2.83l2.5-2.5a2 2 0 012.83 0 .75.75 0 001.06-1.06 3.5 3.5 0 00-4.95 0l-2.5 2.5a3.5 3.5 0 004.95 4.95l1.25-1.25a.75.75 0 00-1.06-1.06l-1.25 1.25a2 2 0 01-2.83 0z"></path></svg></a><span class="notion-h-title">总结</span></span></h3><div class="notion-text notion-block-efabae1e2d3e4e18aa8cf1076c7878ac">简单介绍 OneFlow 模型中如何计算网络的 FLOPs</div><a class="notion-page-link notion-block-ab5d4d9fc9064ff28e32fd0cc98a8c79" href="/ab5d4d9fc9064ff28e32fd0cc98a8c79"></a></main></div>]]></content:encoded>
        </item>
    </channel>
</rss>