{ "source_file": "0930", "source_platform": "知乎专栏", "author": "stone", "title_listed": "语义分割之 Large Kernel Matters", "keywords": "语义分割,大卷积核", "domain": "图像处理", "content_type": "论文精读", "is_paid_flag": "否", "url": "https://zhuanlan.zhihu.com/p/38055224", "publish_date": "2018", "corpus_stage": "应用", "notes": "Large Kernel 语义分割论文解读", "id": "b829e2d460d6", "collected_at": "2026-10-02T18:43:51", "title": "语义分割之Large Kernel Matters - 知乎", "body_markdown": "[![](https://zhuanlan.zhihu.com/p/38055224)](javascript:void\\(0\\))\n![ZhiHu logo](../images/body-e1fa91dbdb48b8055184.png)\n[](https://www.zhihu.com)\n[关注](https://www.zhihu.com/signin?next=%2Ffollow)[推荐](https://www.zhihu.com/signin?next=%2F)[热榜](https://www.zhihu.com/signin?next=%2Fhot)[专栏](https://www.zhihu.com/signin?next=%2Fcolumn-square)[圈子](https://www.zhihu.com/signin?next=%2Fring-feeds)[AI WorksBeta ](https://www.zhihu.com/project-square)[故事](https://www.zhihu.com/fiore/h5/vip-web)\n​\n[直答](https://zhida.zhihu.com/)\n切换模式\n登录/注册\n语义分割之Large Kernel Matters\n首发于[图像语义分割](https://www.zhihu.com/column/c_197474183)\n切换模式\n![语义分割之Large Kernel Matters](../images/000.img)\n# 语义分割之Large Kernel Matters\n[![stone](../images/013.jpg)](https://www.zhihu.com/people/big-big-stone)\n[stone](https://www.zhihu.com/people/big-big-stone)\n收藏之前点个赞呗\n[收录于 · 图像语义分割](https://www.zhihu.com/column/c_197474183)\n27 人赞同了该文章\n## Large Kernel Matters\n## 论文信息\n论文地址：[Large Kernel Matters ——Improve Semantic Segmentation by Global Convolutional Network](https://zhuanlan.zhihu.com/write)\n发表日期：8 Mar 2017\n## 创新点\n 1. 提出全局卷积网络（Global Convolutional Network，GCN），用以同时提高语义分割中分类和定位的准确度。\n 2. 提出Boundary Refinement block（BR）, 用以提高物体边界的定位。\n\n\n## 思想\n在语义分割中，有两个任务，一个是对像素进行分类，一个是对像素进行定位。这两个任务通常是矛盾的：对于分类任务而言，模型需要对各种变换（比如旋转和平移）具有不变形，但是对于定位任务而言，模型又必须对各种变换保持敏感，因为每个像素都需要在正确的位置上进行分类。\n现在很多语义分割方法的主要目标是为了定位，比如下图B所示，这可能会降低分类性能。因此本文提出了GCN来解决以上的矛盾，同时改善分类和定位的性能，如下图C。\n![](../images/004.jpg)\n基于以上的讨论，本文提出了模型的设计思路：\n 1. 从定位角度出发，模型使用FCN来保持位置信息，不使用[全连接层](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E5%85%A8%E8%BF%9E%E6%8E%A5%E5%B1%82&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLlhajov57mjqXlsYIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.orDmodsn-zPC1wItN_QXCQ2bHKMPJrxr1lDvqpwUQms&zhida_source=entity)或者global pooling，因为这些操作会丧失定位信息。\n 2. 从分类角度出发，网络应该使用[大卷积核](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E5%A4%A7%E5%8D%B7%E7%A7%AF%E6%A0%B8&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLlpKfljbfnp6_moLgiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.C4C-UFLHny-qppARPoT0flbIpBH5-GUkZtxcQyWEj-A&zhida_source=entity)，这样能够使分类器具备更强的分类能力来应对各种变换。\n 3. 使用Boundary Refinement block来进行[边界对齐](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E8%BE%B9%E7%95%8C%E5%AF%B9%E9%BD%90&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLovrnnlYzlr7npvZAiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.Sf3Zr613hMTo7dEB2FDwyRHEgJUipppNicYmIDJnuSo&zhida_source=entity)，该模块使用[残差结构](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E6%AE%8B%E5%B7%AE%E7%BB%93%E6%9E%84&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLmrovlt67nu5PmnoQiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.AkhG_p_PKt-z5k_Nu_Wi99D9RSlqY26gjMJ2EOD6hro&zhida_source=entity)。\n\n\n## 模型\n## 1.GCN\n作者发现，目前[state-of-the-art](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=state-of-the-art&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiJzdGF0ZS1vZi10aGUtYXJ0IiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6NzQxMzYxNCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.he90AxQmbyK6tgKeRsOsDMlY5H7ljG9terGE-0GKn_k&zhida_source=entity)的[语义分割](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=4&q=%E8%AF%AD%E4%B9%89%E5%88%86%E5%89%B2&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLor63kuYnliIblibIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjQsInpkX3Rva2VuIjpudWxsfQ.K2Vw38veF88d_9A3l7NyyM98lWLepvFcsFN0510Geeg&zhida_source=entity)方法的设计都是为了更好地进行定位，这在某种程度上可能会降低分类的效果。而分类效果变差可能由于感受野造成的。作者举了个例子，如下图所示，当分类器有效感受野（注意这里的有效感受野不同于[感受野](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=4&q=%E6%84%9F%E5%8F%97%E9%87%8E&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLmhJ_lj5fph44iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjQsInpkX3Rva2VuIjpudWxsfQ.VGt6ycTNQyBsVDhVqIH_9_pfvUL8gzfb9s0FBtjyhWs&zhida_source=entity)。有效感受野的概念来自于论文：Object detectors emerge in deep scene cnns ，这篇论文中称，尽管GoogleNet和ResNet等[深度网络](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E6%B7%B1%E5%BA%A6%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLmt7HluqbnvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.xi2OrwltAs8KxquDKZfsYTTQhRVkx1b73BdBG_61nXw&zhida_source=entity)的感受野通常都大于原图，但网络只能在感受野的一个很小区域获得有效信息，称为有效感受野）足以覆盖整个物体的时候，这时候[分类器](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=3&q=%E5%88%86%E7%B1%BB%E5%99%A8&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLliIbnsbvlmagiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjMsInpkX3Rva2VuIjpudWxsfQ.IhSbWkTz8YHCmJipsqi-eDuc0TqpAJteanYSmxlDAEY&zhida_source=entity)可能可以正确地分类，但是当图像的尺度变大的时候，这个时候有效感受野就会覆盖不了整个物体，这对分类是非常不利的。\n![](../images/005.jpg)\n基于此，作者提出，网络设计要使用FCN，不能使用fully-connected layer或者global pooling，否则丧失位置。另外，卷积的kernel size越大越好。但是越大的卷积核，计算量也越大，所以作者提出使用1xk+kx1和kx1+1xk的卷积组合，这相当于连接了feature map的kxk大小的区域。另外值得注意的是，这里的[卷积](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=6&q=%E5%8D%B7%E7%A7%AF&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLljbfnp68iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjYsInpkX3Rva2VuIjpudWxsfQ.h2iQeEJE4FFdzzaKWCYv2rGaSqVdsyo4N1o7XFctX1M&zhida_source=entity)之后并没有使用[非线性激活函数](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E9%9D%9E%E7%BA%BF%E6%80%A7%E6%BF%80%E6%B4%BB%E5%87%BD%E6%95%B0&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLpnZ7nur_mgKfmv4DmtLvlh73mlbAiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.uUEzFPQ0gkCPpMgHUEZqq7CoIaM-yHwein4CwpQsBQA&zhida_source=entity)。\n![](../images/006.jpg)\n## 2.整体框架\n整体框架如下图所示：\n![](../images/007.jpg)\n从图中可以看到几点：\n 1. 基础网络使用了ResNet作为特征提取，使用FCN作为语义分割的框架。\n 2. 使用了ResNet中不同stage的feature map，因此是多尺度架构。\n 3. GCN模块则用于产生低分辨率的的score map，并上采样与更高分辨率的score map加和产生新的score map。\n 4. 经过最后的上[采样](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=2&q=%E9%87%87%E6%A0%B7&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLph4fmoLciLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjIsInpkX3Rva2VuIjpudWxsfQ.oLbmMD9GcLEVdJiDmkc6FkAW11bXr3oOahMmPWQS7dM&zhida_source=entity)，就输出了预测结果。\n\n\n另外，需要注意的是，在整个网络中还加入了很多Boundary Refinement ，这种结构如下图：\n![](../images/008.jpg)\n可见这是一种[残差连接](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E6%AE%8B%E5%B7%AE%E8%BF%9E%E6%8E%A5&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLmrovlt67ov57mjqUiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.9FpkRg1gSbE-IfB6xYI4BnbjVOAyVJq-tL0v_LwJ1ZE&zhida_source=entity)结构。顶部的wxhx21是一种[粗粒度](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E7%B2%97%E7%B2%92%E5%BA%A6&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLnspfnspLluqYiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.RTxUh5COfyaaRRdRRbBtli7pqxWrEKt5I_IfyPUxB_4&zhida_source=entity)的score map，侧边的残差连接可以对boundary进行refine。两者加和就可以达到Boundary Refinement。\n## 实验\n作者进行了很多实验，这里只列出[PASCAL VOC 2012](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=PASCAL+VOC+2012&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiJQQVNDQUwgVk9DIDIwMTIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.GIh_ivtRxEIgQuwc6Beu6F3eKLrvJmL0gpp8g9mGP_s&zhida_source=entity)的一些可视化结果，更多实验结果请参看原文。\n![](../images/009.jpg)\n编辑于 2018-06-13 13:39\n[深度学习（Deep Learning）](https://www.zhihu.com/topic/19813032)\n[计算机视觉](https://www.zhihu.com/topic/19590195)\n[机器学习](https://www.zhihu.com/topic/19559450)\n​赞同 27​​4 条评论​50 ​1 \n​分享\n​申请转载​\n​\n![](../images/body-7e064b35ba4350074989.jpeg)\n未登录用户\n \n\n4 条评论\n默认\n最新\n[![vvzz](../images/010.jpg)](https://www.zhihu.com/people/54458b7bd33220436e1076769254d816)\n[vvzz](https://www.zhihu.com/people/54458b7bd33220436e1076769254d816)\n这个有开源的代码么\n2019-08-05\n​回复​喜欢\n[![摸头鹰](../images/011.jpg)](https://www.zhihu.com/people/4598605965f440db6a895b9f0a873b18)\n[摸头鹰](https://www.zhihu.com/people/4598605965f440db6a895b9f0a873b18)\n为什么BR加了一个两层卷积的旁路，就说可以对边界进行改善？\n2019-07-24\n​回复​喜欢\n[![stone](../images/012.jpg)](https://www.zhihu.com/people/35ce9d46f0df1f5d30681075645c590d)\n[stone](https://www.zhihu.com/people/35ce9d46f0df1f5d30681075645c590d)\n作者\n我也觉得这个说法没有理由\n2019-07-24\n​回复​喜欢\n展开其他 1 条回复​\n关于作者\n[![stone](../images/013.jpg)](https://www.zhihu.com/people/big-big-stone)\n[stone](https://www.zhihu.com/people/big-big-stone)\n收藏之前点个赞呗\n[ 回答 **30** ](https://www.zhihu.com/people/big-big-stone/answers)[ 文章 **15** ](https://www.zhihu.com/people/big-big-stone/posts)[ 关注者 **10,013** ](https://www.zhihu.com/people/big-big-stone/followers)\n​关注他​发私信\n大家都在搜\n换一换\n[江歌妈妈发长文尘埃终将落定704 万](https://www.zhihu.com/search?q=%E6%B1%9F%E6%AD%8C%E5%A6%88%E5%A6%88%E5%8F%91%E9%95%BF%E6%96%87%E5%B0%98%E5%9F%83%E7%BB%88%E5%B0%86%E8%90%BD%E5%AE%9A&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)热\n[中国男足 0-5 巴勒斯坦562 万](https://www.zhihu.com/search?q=%E4%B8%AD%E5%9B%BD%E7%94%B7%E8%B6%B3+0-5+%E5%B7%B4%E5%8B%92%E6%96%AF%E5%9D%A6&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)热\n[原央视主持人阿丘回应被通报419 万](https://www.zhihu.com/search?q=%E5%8E%9F%E5%A4%AE%E8%A7%86%E4%B8%BB%E6%8C%81%E4%BA%BA%E9%98%BF%E4%B8%98%E5%9B%9E%E5%BA%94%E8%A2%AB%E9%80%9A%E6%8A%A5&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)热\n[华为与赛力斯达成新五年合作409 万](https://www.zhihu.com/search?q=%E5%8D%8E%E4%B8%BA%E4%B8%8E%E8%B5%9B%E5%8A%9B%E6%96%AF%E8%BE%BE%E6%88%90%E6%96%B0%E4%BA%94%E5%B9%B4%E5%90%88%E4%BD%9C&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)热\n[迪拜航空确认航班发生事故398 万](https://www.zhihu.com/search?q=%E8%BF%AA%E6%8B%9C%E8%88%AA%E7%A9%BA%E7%A1%AE%E8%AE%A4%E8%88%AA%E7%8F%AD%E5%8F%91%E7%94%9F%E4%BA%8B%E6%95%85&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)热\n[东航回应网传空姐跪地道歉326 万](https://www.zhihu.com/search?q=%E4%B8%9C%E8%88%AA%E5%9B%9E%E5%BA%94%E7%BD%91%E4%BC%A0%E7%A9%BA%E5%A7%90%E8%B7%AA%E5%9C%B0%E9%81%93%E6%AD%89&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)\n[王楚钦：不知道为什么就是感觉累320 万](https://www.zhihu.com/search?q=%E7%8E%8B%E6%A5%9A%E9%92%A6%EF%BC%9A%E4%B8%8D%E7%9F%A5%E9%81%93%E4%B8%BA%E4%BB%80%E4%B9%88%E5%B0%B1%E6%98%AF%E6%84%9F%E8%A7%89%E7%B4%AF&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)热\n[江苏高考作文《衬衫的价格为 9 磅 15 便士》爆火268 万](https://www.zhihu.com/search?q=%E6%B1%9F%E8%8B%8F%E9%AB%98%E8%80%83%E4%BD%9C%E6%96%87%E3%80%8A%E8%A1%AC%E8%A1%AB%E7%9A%84%E4%BB%B7%E6%A0%BC%E4%B8%BA+9+%E7%A3%85+15+%E4%BE%BF%E5%A3%AB%E3%80%8B%E7%88%86%E7%81%AB&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)\n[女装网店开始用防拆带了265 万](https://www.zhihu.com/search?q=%E5%A5%B3%E8%A3%85%E7%BD%91%E5%BA%97%E5%BC%80%E5%A7%8B%E7%94%A8%E9%98%B2%E6%8B%86%E5%B8%A6%E4%BA%86&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)\n[孕妇骑车别车被司机踹翻257 万](https://www.zhihu.com/search?q=%E5%AD%95%E5%A6%87%E9%AA%91%E8%BD%A6%E5%88%AB%E8%BD%A6%E8%A2%AB%E5%8F%B8%E6%9C%BA%E8%B8%B9%E7%BF%BB&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)\n### 推荐阅读\n# [用简单Mask分类实现语义分割、实例分割“大一统”！Facebook&UIUC开源MaskFormer | 该原创内容首发于极市平台，转载请获得授权并标明出处写在前面在处理语义分割任务和实例分割任务的时候，以前的工作大多采用不同的任务范式。对于语义分割，大多都是作为per-pixel的分类… Fight...发表于论文笔记 ](https://zhuanlan.zhihu.com/p/392731360)[ ![Object Detection from Scratch - A Brief Review](../images/015.jpg)Object Detection from Scratch - A Brief Review 沈志强 ](https://zhuanlan.zhihu.com/p/137410354)# [利用MAE进行语义分割的迁移学习--（2）训练自己的预训练模型MAE 上一篇讲了ViT的一些原理，这一篇我就拿一个实打实的工程给大家做讲解，本文的代码来自github改编的penzhangliang大佬的mae代码.大家直接去github搜索mae第二个就是他的代码，第一个时faceb… 一拳超神 ](https://zhuanlan.zhihu.com/p/480740902)[ ![图文多模态语义融合前的语义对齐——一种单双混合塔多模态模型](../images/017.jpg)图文多模态语义融合前的语义对齐——一种单双混合塔多模态模型 徐土豆发表于计算机视觉... ](https://zhuanlan.zhihu.com/p/462798669) \n\n_想来知乎工作？请发送邮件到 jobs@zhihu.com_\n打开知乎App\n在「我的页」右上角打开扫一扫\n![](../images/body-0f9788d006ab0f4da762.png)\n其他扫码方式：微信\n下载知乎App\n无障碍模式\n验证码登录\n密码登录\n[开通机构号](https://zhuanlan.zhihu.com/org/signup)\n中国 +86\n获取短信验证码\n忘记密码\n登录/注册\n其他方式登录\n未注册手机验证后自动登录，注册即代表同意[《知乎协议》](https://www.zhihu.com/term/zhihu-terms)[《隐私保护指引》](https://www.zhihu.com/term/privacy)\n扫码下载知乎 App\n关闭二维码\n", "metadata": { "title": "语义分割之Large Kernel Matters - 知乎", "description": "Large Kernel Matters论文信息论文地址： Large Kernel Matters ——Improve Semantic Segmentation by Global Convolutional Network发表日期：8 Mar 2017 创新点提出全局卷积网络（Global Convolutional Network…", "keywords": "深度学习（Deep Learning）,计算机视觉,机器学习", "author": null, "og:title": "语义分割之Large Kernel Matters", "og:url": "https://zhuanlan.zhihu.com/p/38055224", "og:description": "Large Kernel Matters论文信息论文地址： Large Kernel Matters ——Improve Semantic Segmentation by Global Convolutional Network发表日期：8 Mar 2017 创新点提出全局卷积网络（Global Convolutional Network…", "og:image": "https://pic1.zhimg.com/v2-5e3b0520ba95f08689ddb086362a81af_720w.jpg?source=172ae18b", "og:type": "article", "og:site_name": "知乎专栏" }, "paywall": { "listed_paid": false, "detected": false, "marker": "", "public_only": false, "policy": "付费/会员内容只保留页面公开可见部分，不绕过付费墙" }, "crawl": { "success": true, "status_code": 200, "error": "" }, "code_blocks": [], "images": [ { "status": "http_403", "url": "https://zhuanlan.zhihu.com/p/38055224", "src": "https://zhuanlan.zhihu.com/p/38055224", "alt": "", "downloaded": false, "discovery": "saved_markdown", "retried_at": "2026-10-20T17:46:45Z" } ], "audio": [], "videos": [], "links": [ { "type": "internal", "href": "https://www.zhihu.com" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2Ffollow" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2F" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2Fhot" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2Fcolumn-square" }, { "type": "internal", "href": "https://www.zhihu.com/signin?next=%2Fring-feeds" }, { "type": "internal", "href": "https://www.zhihu.com/project-square" }, { "type": "internal", "href": "https://www.zhihu.com/fiore/h5/vip-web" }, { "type": "internal", "href": "https://zhida.zhihu.com/" }, { "type": "internal", "href": "https://www.zhihu.com/column/c_197474183" }, { "type": "internal", "href": "https://www.zhihu.com/people/big-big-stone" }, { "type": "internal", "href": "https://zhuanlan.zhihu.com/write" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E5%85%A8%E8%BF%9E%E6%8E%A5%E5%B1%82&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLlhajov57mjqXlsYIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.orDmodsn-zPC1wItN_QXCQ2bHKMPJrxr1lDvqpwUQms&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E5%A4%A7%E5%8D%B7%E7%A7%AF%E6%A0%B8&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLlpKfljbfnp6_moLgiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.C4C-UFLHny-qppARPoT0flbIpBH5-GUkZtxcQyWEj-A&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E8%BE%B9%E7%95%8C%E5%AF%B9%E9%BD%90&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLovrnnlYzlr7npvZAiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.Sf3Zr613hMTo7dEB2FDwyRHEgJUipppNicYmIDJnuSo&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E6%AE%8B%E5%B7%AE%E7%BB%93%E6%9E%84&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLmrovlt67nu5PmnoQiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.AkhG_p_PKt-z5k_Nu_Wi99D9RSlqY26gjMJ2EOD6hro&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=state-of-the-art&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiJzdGF0ZS1vZi10aGUtYXJ0IiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6NzQxMzYxNCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.he90AxQmbyK6tgKeRsOsDMlY5H7ljG9terGE-0GKn_k&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=4&q=%E8%AF%AD%E4%B9%89%E5%88%86%E5%89%B2&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLor63kuYnliIblibIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjQsInpkX3Rva2VuIjpudWxsfQ.K2Vw38veF88d_9A3l7NyyM98lWLepvFcsFN0510Geeg&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=4&q=%E6%84%9F%E5%8F%97%E9%87%8E&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLmhJ_lj5fph44iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjQsInpkX3Rva2VuIjpudWxsfQ.VGt6ycTNQyBsVDhVqIH_9_pfvUL8gzfb9s0FBtjyhWs&zhida_source=entity" }, { "type": "internal", "href": "https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E6%B7%B1%E5%BA%A6%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLmt7HluqbnvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.xi2OrwltAs8KxquDKZfsYTTQhRVkx1b73BdBG_61nXw&zhida_source=entity" }, { "type": "external", "href": "javascript:void(0)" } ], "repair_version": 2, "repaired_at": "2026-10-20T17:26:13Z", "source_url": "https://zhuanlan.zhihu.com/p/38055224", "delivery_status": "unverified", "discovery_status": "http_403", "missing_reason": "源页面未重新核验:http_403；已按保留正文补采图片", "repair_error": "", "actual_video_files": [], "expected_saved_body_images": 1, "image_discovery": "saved_markdown", "source_body_markdown": "[![](https://zhuanlan.zhihu.com/p/38055224)](javascript:void\\(0\\))\n![ZhiHu logo](https://static.zhihu.com/heifetz/assets/wechat-share-logo.39ea9ecd.png)\n[](https://www.zhihu.com)\n[关注](https://www.zhihu.com/signin?next=%2Ffollow)[推荐](https://www.zhihu.com/signin?next=%2F)[热榜](https://www.zhihu.com/signin?next=%2Fhot)[专栏](https://www.zhihu.com/signin?next=%2Fcolumn-square)[圈子](https://www.zhihu.com/signin?next=%2Fring-feeds)[AI WorksBeta ](https://www.zhihu.com/project-square)[故事](https://www.zhihu.com/fiore/h5/vip-web)\n​\n[直答](https://zhida.zhihu.com/)\n切换模式\n登录/注册\n语义分割之Large Kernel Matters\n首发于[图像语义分割](https://www.zhihu.com/column/c_197474183)\n切换模式\n![语义分割之Large Kernel Matters](https://pic1.zhimg.com/70/v2-5e3b0520ba95f08689ddb086362a81af_1440w.image?source=172ae18b&biz_tag=Post)\n# 语义分割之Large Kernel Matters\n[![stone](https://pic1.zhimg.com/36f310418_l.jpg?source=32738c0c&needBackground=1)](https://www.zhihu.com/people/big-big-stone)\n[stone](https://www.zhihu.com/people/big-big-stone)\n收藏之前点个赞呗\n[收录于 · 图像语义分割](https://www.zhihu.com/column/c_197474183)\n27 人赞同了该文章\n## Large Kernel Matters\n## 论文信息\n论文地址：[Large Kernel Matters ——Improve Semantic Segmentation by Global Convolutional Network](https://zhuanlan.zhihu.com/write)\n发表日期：8 Mar 2017\n## 创新点\n 1. 提出全局卷积网络（Global Convolutional Network，GCN），用以同时提高语义分割中分类和定位的准确度。\n 2. 提出Boundary Refinement block（BR）, 用以提高物体边界的定位。\n\n\n## 思想\n在语义分割中，有两个任务，一个是对像素进行分类，一个是对像素进行定位。这两个任务通常是矛盾的：对于分类任务而言，模型需要对各种变换（比如旋转和平移）具有不变形，但是对于定位任务而言，模型又必须对各种变换保持敏感，因为每个像素都需要在正确的位置上进行分类。\n现在很多语义分割方法的主要目标是为了定位，比如下图B所示，这可能会降低分类性能。因此本文提出了GCN来解决以上的矛盾，同时改善分类和定位的性能，如下图C。\n![](https://pic3.zhimg.com/v2-d1378d80ffb72320711ca1a9a4865e04_1440w.jpg)\n基于以上的讨论，本文提出了模型的设计思路：\n 1. 从定位角度出发，模型使用FCN来保持位置信息，不使用[全连接层](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E5%85%A8%E8%BF%9E%E6%8E%A5%E5%B1%82&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLlhajov57mjqXlsYIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.orDmodsn-zPC1wItN_QXCQ2bHKMPJrxr1lDvqpwUQms&zhida_source=entity)或者global pooling，因为这些操作会丧失定位信息。\n 2. 从分类角度出发，网络应该使用[大卷积核](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E5%A4%A7%E5%8D%B7%E7%A7%AF%E6%A0%B8&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLlpKfljbfnp6_moLgiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.C4C-UFLHny-qppARPoT0flbIpBH5-GUkZtxcQyWEj-A&zhida_source=entity)，这样能够使分类器具备更强的分类能力来应对各种变换。\n 3. 使用Boundary Refinement block来进行[边界对齐](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E8%BE%B9%E7%95%8C%E5%AF%B9%E9%BD%90&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLovrnnlYzlr7npvZAiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.Sf3Zr613hMTo7dEB2FDwyRHEgJUipppNicYmIDJnuSo&zhida_source=entity)，该模块使用[残差结构](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E6%AE%8B%E5%B7%AE%E7%BB%93%E6%9E%84&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLmrovlt67nu5PmnoQiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.AkhG_p_PKt-z5k_Nu_Wi99D9RSlqY26gjMJ2EOD6hro&zhida_source=entity)。\n\n\n## 模型\n## 1.GCN\n作者发现，目前[state-of-the-art](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=state-of-the-art&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiJzdGF0ZS1vZi10aGUtYXJ0IiwiemhpZGFfc291cmNlIjoiZW50aXR5IiwiY29udGVudF9pZCI6NzQxMzYxNCwiY29udGVudF90eXBlIjoiQXJ0aWNsZSIsIm1hdGNoX29yZGVyIjoxLCJ6ZF90b2tlbiI6bnVsbH0.he90AxQmbyK6tgKeRsOsDMlY5H7ljG9terGE-0GKn_k&zhida_source=entity)的[语义分割](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=4&q=%E8%AF%AD%E4%B9%89%E5%88%86%E5%89%B2&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLor63kuYnliIblibIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjQsInpkX3Rva2VuIjpudWxsfQ.K2Vw38veF88d_9A3l7NyyM98lWLepvFcsFN0510Geeg&zhida_source=entity)方法的设计都是为了更好地进行定位，这在某种程度上可能会降低分类的效果。而分类效果变差可能由于感受野造成的。作者举了个例子，如下图所示，当分类器有效感受野（注意这里的有效感受野不同于[感受野](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=4&q=%E6%84%9F%E5%8F%97%E9%87%8E&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLmhJ_lj5fph44iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjQsInpkX3Rva2VuIjpudWxsfQ.VGt6ycTNQyBsVDhVqIH_9_pfvUL8gzfb9s0FBtjyhWs&zhida_source=entity)。有效感受野的概念来自于论文：Object detectors emerge in deep scene cnns ，这篇论文中称，尽管GoogleNet和ResNet等[深度网络](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E6%B7%B1%E5%BA%A6%E7%BD%91%E7%BB%9C&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLmt7HluqbnvZHnu5wiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.xi2OrwltAs8KxquDKZfsYTTQhRVkx1b73BdBG_61nXw&zhida_source=entity)的感受野通常都大于原图，但网络只能在感受野的一个很小区域获得有效信息，称为有效感受野）足以覆盖整个物体的时候，这时候[分类器](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=3&q=%E5%88%86%E7%B1%BB%E5%99%A8&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLliIbnsbvlmagiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjMsInpkX3Rva2VuIjpudWxsfQ.IhSbWkTz8YHCmJipsqi-eDuc0TqpAJteanYSmxlDAEY&zhida_source=entity)可能可以正确地分类，但是当图像的尺度变大的时候，这个时候有效感受野就会覆盖不了整个物体，这对分类是非常不利的。\n![](https://pic2.zhimg.com/v2-f8f391563b013bb3d66afe75050119e9_1440w.jpg)\n基于此，作者提出，网络设计要使用FCN，不能使用fully-connected layer或者global pooling，否则丧失位置。另外，卷积的kernel size越大越好。但是越大的卷积核，计算量也越大，所以作者提出使用1xk+kx1和kx1+1xk的卷积组合，这相当于连接了feature map的kxk大小的区域。另外值得注意的是，这里的[卷积](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=6&q=%E5%8D%B7%E7%A7%AF&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLljbfnp68iLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjYsInpkX3Rva2VuIjpudWxsfQ.h2iQeEJE4FFdzzaKWCYv2rGaSqVdsyo4N1o7XFctX1M&zhida_source=entity)之后并没有使用[非线性激活函数](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E9%9D%9E%E7%BA%BF%E6%80%A7%E6%BF%80%E6%B4%BB%E5%87%BD%E6%95%B0&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLpnZ7nur_mgKfmv4DmtLvlh73mlbAiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.uUEzFPQ0gkCPpMgHUEZqq7CoIaM-yHwein4CwpQsBQA&zhida_source=entity)。\n![](https://picx.zhimg.com/v2-87f92fda5d3a9c7bd1df2617b902cadd_1440w.jpg)\n## 2.整体框架\n整体框架如下图所示：\n![](https://pica.zhimg.com/v2-686b7584800622bd5e3626f03e14b6ee_1440w.jpg)\n从图中可以看到几点：\n 1. 基础网络使用了ResNet作为特征提取，使用FCN作为语义分割的框架。\n 2. 使用了ResNet中不同stage的feature map，因此是多尺度架构。\n 3. GCN模块则用于产生低分辨率的的score map，并上采样与更高分辨率的score map加和产生新的score map。\n 4. 经过最后的上[采样](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=2&q=%E9%87%87%E6%A0%B7&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLph4fmoLciLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjIsInpkX3Rva2VuIjpudWxsfQ.oLbmMD9GcLEVdJiDmkc6FkAW11bXr3oOahMmPWQS7dM&zhida_source=entity)，就输出了预测结果。\n\n\n另外，需要注意的是，在整个网络中还加入了很多Boundary Refinement ，这种结构如下图：\n![](https://pica.zhimg.com/v2-00a791b2301ae2aba99e8e9cfa2ba34c_1440w.jpg)\n可见这是一种[残差连接](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E6%AE%8B%E5%B7%AE%E8%BF%9E%E6%8E%A5&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLmrovlt67ov57mjqUiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.9FpkRg1gSbE-IfB6xYI4BnbjVOAyVJq-tL0v_LwJ1ZE&zhida_source=entity)结构。顶部的wxhx21是一种[粗粒度](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=%E7%B2%97%E7%B2%92%E5%BA%A6&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiLnspfnspLluqYiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.RTxUh5COfyaaRRdRRbBtli7pqxWrEKt5I_IfyPUxB_4&zhida_source=entity)的score map，侧边的残差连接可以对boundary进行refine。两者加和就可以达到Boundary Refinement。\n## 实验\n作者进行了很多实验，这里只列出[PASCAL VOC 2012](https://zhida.zhihu.com/search?content_id=7413614&content_type=Article&match_order=1&q=PASCAL+VOC+2012&zd_token=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJ6aGlkYV9zZXJ2ZXIiLCJleHAiOjE3OTExMzk0NDAsInEiOiJQQVNDQUwgVk9DIDIwMTIiLCJ6aGlkYV9zb3VyY2UiOiJlbnRpdHkiLCJjb250ZW50X2lkIjo3NDEzNjE0LCJjb250ZW50X3R5cGUiOiJBcnRpY2xlIiwibWF0Y2hfb3JkZXIiOjEsInpkX3Rva2VuIjpudWxsfQ.GIh_ivtRxEIgQuwc6Beu6F3eKLrvJmL0gpp8g9mGP_s&zhida_source=entity)的一些可视化结果，更多实验结果请参看原文。\n![](https://pic1.zhimg.com/v2-c00341190ba0f3af4833ff965d8c8028_1440w.jpg)\n编辑于 2018-06-13 13:39\n[深度学习（Deep Learning）](https://www.zhihu.com/topic/19813032)\n[计算机视觉](https://www.zhihu.com/topic/19590195)\n[机器学习](https://www.zhihu.com/topic/19559450)\n​赞同 27​​4 条评论​50 ​1 \n​分享\n​申请转载​\n​\n![](https://pic1.zhimg.com/v2-d41c2ceaed8f51999522f903672a521f_l.jpeg)\n未登录用户\n \n\n4 条评论\n默认\n最新\n[![vvzz](https://pic1.zhimg.com/33096da7ae4c2d01df09f6df18b021f6_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/54458b7bd33220436e1076769254d816)\n[vvzz](https://www.zhihu.com/people/54458b7bd33220436e1076769254d816)\n这个有开源的代码么\n2019-08-05\n​回复​喜欢\n[![摸头鹰](https://picx.zhimg.com/v2-60a6d744980efe82b57d6f6db8f503f1_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/4598605965f440db6a895b9f0a873b18)\n[摸头鹰](https://www.zhihu.com/people/4598605965f440db6a895b9f0a873b18)\n为什么BR加了一个两层卷积的旁路，就说可以对边界进行改善？\n2019-07-24\n​回复​喜欢\n[![stone](https://picx.zhimg.com/36f310418_l.jpg?source=06d4cd63)](https://www.zhihu.com/people/35ce9d46f0df1f5d30681075645c590d)\n[stone](https://www.zhihu.com/people/35ce9d46f0df1f5d30681075645c590d)\n作者\n我也觉得这个说法没有理由\n2019-07-24\n​回复​喜欢\n展开其他 1 条回复​\n关于作者\n[![stone](https://pic1.zhimg.com/36f310418_l.jpg?source=32738c0c&needBackground=1)](https://www.zhihu.com/people/big-big-stone)\n[stone](https://www.zhihu.com/people/big-big-stone)\n收藏之前点个赞呗\n[ 回答 **30** ](https://www.zhihu.com/people/big-big-stone/answers)[ 文章 **15** ](https://www.zhihu.com/people/big-big-stone/posts)[ 关注者 **10,013** ](https://www.zhihu.com/people/big-big-stone/followers)\n​关注他​发私信\n大家都在搜\n换一换\n[江歌妈妈发长文尘埃终将落定704 万](https://www.zhihu.com/search?q=%E6%B1%9F%E6%AD%8C%E5%A6%88%E5%A6%88%E5%8F%91%E9%95%BF%E6%96%87%E5%B0%98%E5%9F%83%E7%BB%88%E5%B0%86%E8%90%BD%E5%AE%9A&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)热\n[中国男足 0-5 巴勒斯坦562 万](https://www.zhihu.com/search?q=%E4%B8%AD%E5%9B%BD%E7%94%B7%E8%B6%B3+0-5+%E5%B7%B4%E5%8B%92%E6%96%AF%E5%9D%A6&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)热\n[原央视主持人阿丘回应被通报419 万](https://www.zhihu.com/search?q=%E5%8E%9F%E5%A4%AE%E8%A7%86%E4%B8%BB%E6%8C%81%E4%BA%BA%E9%98%BF%E4%B8%98%E5%9B%9E%E5%BA%94%E8%A2%AB%E9%80%9A%E6%8A%A5&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)热\n[华为与赛力斯达成新五年合作409 万](https://www.zhihu.com/search?q=%E5%8D%8E%E4%B8%BA%E4%B8%8E%E8%B5%9B%E5%8A%9B%E6%96%AF%E8%BE%BE%E6%88%90%E6%96%B0%E4%BA%94%E5%B9%B4%E5%90%88%E4%BD%9C&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)热\n[迪拜航空确认航班发生事故398 万](https://www.zhihu.com/search?q=%E8%BF%AA%E6%8B%9C%E8%88%AA%E7%A9%BA%E7%A1%AE%E8%AE%A4%E8%88%AA%E7%8F%AD%E5%8F%91%E7%94%9F%E4%BA%8B%E6%95%85&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)热\n[东航回应网传空姐跪地道歉326 万](https://www.zhihu.com/search?q=%E4%B8%9C%E8%88%AA%E5%9B%9E%E5%BA%94%E7%BD%91%E4%BC%A0%E7%A9%BA%E5%A7%90%E8%B7%AA%E5%9C%B0%E9%81%93%E6%AD%89&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)\n[王楚钦：不知道为什么就是感觉累320 万](https://www.zhihu.com/search?q=%E7%8E%8B%E6%A5%9A%E9%92%A6%EF%BC%9A%E4%B8%8D%E7%9F%A5%E9%81%93%E4%B8%BA%E4%BB%80%E4%B9%88%E5%B0%B1%E6%98%AF%E6%84%9F%E8%A7%89%E7%B4%AF&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)热\n[江苏高考作文《衬衫的价格为 9 磅 15 便士》爆火268 万](https://www.zhihu.com/search?q=%E6%B1%9F%E8%8B%8F%E9%AB%98%E8%80%83%E4%BD%9C%E6%96%87%E3%80%8A%E8%A1%AC%E8%A1%AB%E7%9A%84%E4%BB%B7%E6%A0%BC%E4%B8%BA+9+%E7%A3%85+15+%E4%BE%BF%E5%A3%AB%E3%80%8B%E7%88%86%E7%81%AB&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)\n[女装网店开始用防拆带了265 万](https://www.zhihu.com/search?q=%E5%A5%B3%E8%A3%85%E7%BD%91%E5%BA%97%E5%BC%80%E5%A7%8B%E7%94%A8%E9%98%B2%E6%8B%86%E5%B8%A6%E4%BA%86&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)\n[孕妇骑车别车被司机踹翻257 万](https://www.zhihu.com/search?q=%E5%AD%95%E5%A6%87%E9%AA%91%E8%BD%A6%E5%88%AB%E8%BD%A6%E8%A2%AB%E5%8F%B8%E6%9C%BA%E8%B8%B9%E7%BF%BB&search_source=Trending&utm_content=search_hot&utm_medium=organic&utm_source=zhihu&type=content&request_click_id=a1e76c10-fdb1-4a1b-93e1-ac89fa088185)\n### 推荐阅读\n# [用简单Mask分类实现语义分割、实例分割“大一统”！Facebook&UIUC开源MaskFormer | 该原创内容首发于极市平台，转载请获得授权并标明出处写在前面在处理语义分割任务和实例分割任务的时候，以前的工作大多采用不同的任务范式。对于语义分割，大多都是作为per-pixel的分类… Fight...发表于论文笔记 ](https://zhuanlan.zhihu.com/p/392731360)[ ![Object Detection from Scratch - A Brief Review](https://picx.zhimg.com/v2-2c043d8d9050d143bf44406b6470de24_250x0.jpg?source=172ae18b)Object Detection from Scratch - A Brief Review 沈志强 ](https://zhuanlan.zhihu.com/p/137410354)# [利用MAE进行语义分割的迁移学习--（2）训练自己的预训练模型MAE 上一篇讲了ViT的一些原理，这一篇我就拿一个实打实的工程给大家做讲解，本文的代码来自github改编的penzhangliang大佬的mae代码.大家直接去github搜索mae第二个就是他的代码，第一个时faceb… 一拳超神 ](https://zhuanlan.zhihu.com/p/480740902)[ ![图文多模态语义融合前的语义对齐——一种单双混合塔多模态模型](https://picx.zhimg.com/v2-a42f7dc54c71a81d06b9ec828cb77019_250x0.jpg?source=172ae18b)图文多模态语义融合前的语义对齐——一种单双混合塔多模态模型 徐土豆发表于计算机视觉... ](https://zhuanlan.zhihu.com/p/462798669) \n\n_想来知乎工作？请发送邮件到 jobs@zhihu.com_\n打开知乎App\n在「我的页」右上角打开扫一扫\n![](https://picx.zhimg.com/v2-9e41ea16bdfbe9cf4896617ecad5b4ca.png)\n其他扫码方式：微信\n下载知乎App\n无障碍模式\n验证码登录\n密码登录\n[开通机构号](https://zhuanlan.zhihu.com/org/signup)\n中国 +86\n获取短信验证码\n忘记密码\n登录/注册\n其他方式登录\n未注册手机验证后自动登录，注册即代表同意[《知乎协议》](https://www.zhihu.com/term/zhihu-terms)[《隐私保护指引》](https://www.zhihu.com/term/privacy)\n扫码下载知乎 App\n关闭二维码\n", "body_base_path": "attachments/b829e2d460d6/pages/", "body_markdown_path": "attachments/b829e2d460d6/pages/archived-body.md", "local_html": "attachments/b829e2d460d6/pages/archived-body.html", "body_source": "prior_capture" }