{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"R-CNN：让已经会看图的网络看一个区域","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html","markdown":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.md","pages":[{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","unit":"多个区域怎样复用同一份视觉计算？","unitId":"region-computation-reuse","chapter":"II","number":1}],"aliases":[{"oldId":"explanation","targetId":"region-pretrained-computation","operation":"split-explanation-range"},{"oldId":"b-db706491d73a185536b1641d","targetId":"b-db706491d73a185536b1641d","page":"region-pretrained-computation","operation":"move","beforeSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","afterSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","reason":"Same source body in authored semantic page"},{"oldId":"b-df15c67fbe08f1489f454fa0","targetId":"b-df15c67fbe08f1489f454fa0","page":"region-pretrained-computation","operation":"move","beforeSHA256":"bc51d883c7be62b6cbf1f7b7311a59b087ca23b5bf3285d2396d1063d903fa99","afterSHA256":"bc51d883c7be62b6cbf1f7b7311a59b087ca23b5bf3285d2396d1063d903fa99","reason":"Same source body in authored semantic page"},{"oldId":"b-9ee9c9a8113e4bb8ba354ff1","targetId":"b-9ee9c9a8113e4bb8ba354ff1","page":"region-pretrained-computation","operation":"move","beforeSHA256":"f5e46123e517748cb0b2feb64ecff80b4cc4c546f33d59cc42356db94db6d2c9","afterSHA256":"f5e46123e517748cb0b2feb64ecff80b4cc4c546f33d59cc42356db94db6d2c9","reason":"Same source body in authored semantic page"},{"oldId":"b-a7d8269d387b6e2a1a41998c","targetId":"b-a7d8269d387b6e2a1a41998c","page":"region-pretrained-computation","operation":"rewrite","beforeSHA256":"168961522b4a504243e32c0559e53994d1169dd7de5daac65759fb344e6cd140","afterSHA256":"44e2f2002a3d094df48052d148682fc77608e6148c8194cfc4d310812c1796df","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-014db349e2cc630e843149a2","targetId":"b-014db349e2cc630e843149a2","page":"region-pretrained-computation","operation":"move","beforeSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","afterSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","reason":"Same source body in authored semantic page"},{"oldId":"b-a8495dd258f341f95c1df3c3","targetId":"b-a8495dd258f341f95c1df3c3","page":"region-pretrained-computation","operation":"move","beforeSHA256":"4ddc2eaf1803d689abad356c49fe8ae2f13d816b2462dac02bd23f3a9dd061d0","afterSHA256":"4ddc2eaf1803d689abad356c49fe8ae2f13d816b2462dac02bd23f3a9dd061d0","reason":"Same source body in authored semantic page"},{"oldId":"b-970e13401b12f967a4320c87","targetId":"b-970e13401b12f967a4320c87","page":"region-pretrained-computation","operation":"move","beforeSHA256":"23536efc2cc3debb4fb2ccea58781bc8a4934a6118c383dbe174606efaaf65b8","afterSHA256":"23536efc2cc3debb4fb2ccea58781bc8a4934a6118c383dbe174606efaaf65b8","reason":"Same source body in authored semantic page"},{"oldId":"b-1c6f61f68f50309e7d67811d","targetId":"b-1c6f61f68f50309e7d67811d","page":"region-pretrained-computation","operation":"move","beforeSHA256":"a620c6e0f57f9d21a737c4af613fe460cac7e0d69e0cf0248c6827ec83e987db","afterSHA256":"a620c6e0f57f9d21a737c4af613fe460cac7e0d69e0cf0248c6827ec83e987db","reason":"Same source body in authored semantic page"},{"oldId":"b-b9cd24812bfead847a6ba0a5","targetId":"b-b9cd24812bfead847a6ba0a5","page":"region-pretrained-computation","operation":"move","beforeSHA256":"861774ba5faaffdebcc9f839c46f4495cc31d8be666790b3138545d99ce00d44","afterSHA256":"861774ba5faaffdebcc9f839c46f4495cc31d8be666790b3138545d99ce00d44","reason":"Same source body in authored semantic page"},{"oldId":"b-651424ae6acac0a4ea30ad02","targetId":"b-651424ae6acac0a4ea30ad02","page":"region-pretrained-computation","operation":"move","beforeSHA256":"178bc2569407866ae7371a893269c1c7e57bc6d16f7f1c044f037e1730c1062d","afterSHA256":"178bc2569407866ae7371a893269c1c7e57bc6d16f7f1c044f037e1730c1062d","reason":"Same source body in authored semantic page"},{"oldId":"b-a041b1b8048c190f1d7766ed","targetId":"b-a041b1b8048c190f1d7766ed","page":"region-pretrained-computation","operation":"move","beforeSHA256":"f16be0424762a44344338696f06ef266c4571d15367ae2cbbd28fac98ac7e9cb","afterSHA256":"f16be0424762a44344338696f06ef266c4571d15367ae2cbbd28fac98ac7e9cb","reason":"Same source body in authored semantic page"},{"oldId":"b-9af0333ffb0e4a95b064bad5","targetId":"b-9af0333ffb0e4a95b064bad5","page":"region-pretrained-computation","operation":"move","beforeSHA256":"b25685f04cd1b885f3a5999e6a912460c7424021cdc5b2190584242086a82681","afterSHA256":"b25685f04cd1b885f3a5999e6a912460c7424021cdc5b2190584242086a82681","reason":"Same source body in authored semantic page"},{"oldId":"b-5039e659d4f8286848729331","targetId":"b-5039e659d4f8286848729331","page":"region-pretrained-computation","operation":"move","beforeSHA256":"f7230090f4acc46b2a1ac55d497d3030034f199ad1e9d33a7f04738dcf047b9d","afterSHA256":"f7230090f4acc46b2a1ac55d497d3030034f199ad1e9d33a7f04738dcf047b9d","reason":"Same source body in authored semantic page"},{"oldId":"b-e24d120a6aed0cb7d164a0af","targetId":"b-e24d120a6aed0cb7d164a0af","page":"region-pretrained-computation","operation":"move","beforeSHA256":"690ddccb15578a5c7e7b59a35026f1eef0293bd8670dccd0d6f6f600414f7eb4","afterSHA256":"690ddccb15578a5c7e7b59a35026f1eef0293bd8670dccd0d6f6f600414f7eb4","reason":"Same source body in authored semantic page"},{"oldId":"b-4ec73e31d68009582ef971cd","targetId":"b-4ec73e31d68009582ef971cd","page":"region-pretrained-computation","operation":"move","beforeSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","afterSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","reason":"Same source body in authored semantic page"},{"oldId":"b-0bc10a6781f68be89bedfc7d","targetId":"b-0bc10a6781f68be89bedfc7d","page":"region-pretrained-computation","operation":"move","beforeSHA256":"a3163e9b7618b50175088fce50e8c02bf38584dbcb466052e3f6745aa91110a4","afterSHA256":"a3163e9b7618b50175088fce50e8c02bf38584dbcb466052e3f6745aa91110a4","reason":"Same source body in authored semantic page"},{"oldId":"b-4cd19b1fc6b80c98ae7d3dc7","targetId":"b-4cd19b1fc6b80c98ae7d3dc7","page":"region-pretrained-computation","operation":"move","beforeSHA256":"2e0b8ea011e1741562652d0865d8840fed4d61e7f8b1d1899b73400f0e06f2f0","afterSHA256":"2e0b8ea011e1741562652d0865d8840fed4d61e7f8b1d1899b73400f0e06f2f0","reason":"Same source body in authored semantic page"},{"oldId":"b-c2e93e65ad685d8a81518390","targetId":"b-c2e93e65ad685d8a81518390","page":"region-pretrained-computation","operation":"move","beforeSHA256":"b1296df119df9b111607b73a175dc1db15fc7678a5b61de3cb114f4f3711ec4c","afterSHA256":"b1296df119df9b111607b73a175dc1db15fc7678a5b61de3cb114f4f3711ec4c","reason":"Same source body in authored semantic page"},{"oldId":"b-97305282ac4df7da0d07b8f5","targetId":"b-97305282ac4df7da0d07b8f5","page":"region-pretrained-computation","operation":"move","beforeSHA256":"b1296df119df9b111607b73a175dc1db15fc7678a5b61de3cb114f4f3711ec4c","afterSHA256":"b1296df119df9b111607b73a175dc1db15fc7678a5b61de3cb114f4f3711ec4c","reason":"Same source body in authored semantic page"},{"oldId":"region-pretrained-computation-title","targetId":"b-fe8a81774236afb24f315206","operation":"retain-source-identity"}],"blocks":[{"id":"b-fe8a81774236afb24f315206","type":"h2","scope":"region-pretrained-computation","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-fe8a81774236afb24f315206","content":"## 让预训练网络读取一个区域","sha256":"003b357e456b922c96de268b06f55b7624473d730b6817e3373661b4f0ffc8ba","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-db706491d73a185536b1641d","prerequisites":[{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]},{"id":"b-db706491d73a185536b1641d","type":"h2","scope":"explanation","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-db706491d73a185536b1641d","content":"## 解释、算例与证据","sha256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-fe8a81774236afb24f315206","nextBlock":"b-df15c67fbe08f1489f454fa0","prerequisites":[{"id":"b-fe8a81774236afb24f315206","url":"#b-fe8a81774236afb24f315206","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]},{"id":"b-df15c67fbe08f1489f454fa0","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-df15c67fbe08f1489f454fa0","content":"# P27 · R-CNN：让已经会看图的网络看一个区域","sha256":"bc51d883c7be62b6cbf1f7b7311a59b087ca23b5bf3285d2396d1063d903fa99","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-db706491d73a185536b1641d","nextBlock":"b-9ee9c9a8113e4bb8ba354ff1","prerequisites":[{"id":"b-db706491d73a185536b1641d","url":"#b-db706491d73a185536b1641d","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]},{"id":"b-9ee9c9a8113e4bb8ba354ff1","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-9ee9c9a8113e4bb8ba354ff1","content":"历史上 R-CNN 将监督分类预训练用于区域检测，支持了“先学表示，再适配新任务”的正面判断。按区域提取特征、学习前景/背景和位置修正三个动作都需要具体输入。它不是 MoCo 或 MAE 的历史后继。","sha256":"f5e46123e517748cb0b2feb64ecff80b4cc4c546f33d59cc42356db94db6d2c9","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-df15c67fbe08f1489f454fa0","nextBlock":"b-a7d8269d387b6e2a1a41998c","prerequisites":[{"id":"b-df15c67fbe08f1489f454fa0","url":"#b-df15c67fbe08f1489f454fa0","role":"section-title"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]},{"id":"b-a7d8269d387b6e2a1a41998c","type":"figure","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-a7d8269d387b6e2a1a41998c","content":"图内标签与数值：Lecture 03 课堂图解 P27；候选区域 → 裁出并缩放；红杯候选；预训练 CNN；ImageNet 分类权重；区域类别：杯子；位置：相对框修正；Girshick · Donahue · Darrell · Malik；迁移参数，还要学习检测任务；R-CNN · 2013 预印本 / CVPR 2014。课堂顺序不是历史发生顺序。\n\n同一场景的候选框、红杯裁剪、ImageNet 预训练 CNN 与区域类别和框修正。\n\n- 候选区域 → 裁出并缩放\n- 红杯候选\n- 预训练 CNN\n- ImageNet 分类权重\n- 区域类别：杯子\n- 位置：相对框修正\n- Girshick · Donahue · Darrell · Malik\n- 迁移参数，还要学习检测任务\n- R-CNN · 2013 预印本 / CVPR 2014。课堂顺序不是历史发生顺序。同一场景的候选框、红杯裁剪、ImageNet 预训练 CNN 与区域类别和框修正。 [查看对应图解状态](/#/l3-r2-p27/1)[查看完整图解](/course/figures/l3-r2-p27-1.html)","sha256":"44e2f2002a3d094df48052d148682fc77608e6148c8194cfc4d310812c1796df","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"同一场景的候选框、红杯裁剪、ImageNet 预训练 CNN 与区域类别和框修正。 查看对应图解状态","labels":["候选区域 → 裁出并缩放","红杯候选","预训练 CNN","ImageNet 分类权重","区域类别：杯子","位置：相对框修正","Girshick · Donahue · Darrell · Malik","迁移参数，还要学习检测任务","R-CNN · 2013 预印本 / CVPR 2014。课堂顺序不是历史发生顺序。"],"description":"同一场景的候选框、红杯裁剪、ImageNet 预训练 CNN 与区域类别和框修正。候选区域 → 裁出并缩放红杯候选预训练 CNNImageNet 分类权重区域类别：杯子位置：相对框修正Girshick · Donahue · Darrell · Malik迁移参数，还要学习检测任务R-CNN · 2013 预印本 / CVPR 2014。课堂顺序不是历史发生顺序。","evidenceIdentity":"同一场景的候选框、红杯裁剪、ImageNet 预训练 CNN 与区域类别和框修正。 查看对应图解状态","images":[{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""},{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""}]},"previousBlock":"b-9ee9c9a8113e4bb8ba354ff1","nextBlock":"b-014db349e2cc630e843149a2","prerequisites":[{"id":"b-df15c67fbe08f1489f454fa0","url":"#b-df15c67fbe08f1489f454fa0","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-014db349e2cc630e843149a2","type":"h2","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-014db349e2cc630e843149a2","content":"## 课堂讲述","sha256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-a7d8269d387b6e2a1a41998c","nextBlock":"b-a8495dd258f341f95c1df3c3","prerequisites":[{"id":"b-df15c67fbe08f1489f454fa0","url":"#b-df15c67fbe08f1489f454fa0","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]},{"id":"b-a8495dd258f341f95c1df3c3","type":"p","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-a8495dd258f341f95c1df3c3","content":"R-CNN 的重要启发，是把分类预训练得到的视觉计算用于检测。先提出一批可能包含物体的区域，把区域裁出来，送进网络，再判断类别和位置。\n这里迁移的是视觉网络及其参数，不是 ImageNet 原来的答案列表。它还要面向检测作适配，学习前景、背景以及定位任务。这个结果让“在大数据上先学，再用到别的任务”成为很具体的做法。\n原始 R-CNN 是 Ross Girshick、Jeff Donahue、Trevor Darrell 和 Jitendra Malik 的工作。今天我们从自监督接到它，是在整理表示与任务的关系；它本来的重要背景是分类预训练，不要把历史改写成它先使用了 MoCo 或 MAE。","sha256":"4ddc2eaf1803d689abad356c49fe8ae2f13d816b2462dac02bd23f3a9dd061d0","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-014db349e2cc630e843149a2","nextBlock":"b-970e13401b12f967a4320c87","prerequisites":[{"id":"b-014db349e2cc630e843149a2","url":"#b-014db349e2cc630e843149a2","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]},{"id":"b-970e13401b12f967a4320c87","type":"h3","scope":"rcnn-transfer-pipeline","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-970e13401b12f967a4320c87","content":"### R-CNN 迁移的是已学会的视觉计算","sha256":"23536efc2cc3debb4fb2ccea58781bc8a4934a6118c383dbe174606efaaf65b8","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"R-CNN 迁移的是已学会的视觉计算","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":"rcnn-transfer-pipeline","formulas":[],"previousBlock":"b-a8495dd258f341f95c1df3c3","nextBlock":"b-1c6f61f68f50309e7d67811d","prerequisites":[{"id":"b-014db349e2cc630e843149a2","url":"#b-014db349e2cc630e843149a2","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"},{"id":"b-1c6f61f68f50309e7d67811d","url":"#b-1c6f61f68f50309e7d67811d","relation":"same-reasoning"},{"id":"b-b9cd24812bfead847a6ba0a5","url":"#b-b9cd24812bfead847a6ba0a5","relation":"same-reasoning"},{"id":"b-651424ae6acac0a4ea30ad02","url":"#b-651424ae6acac0a4ea30ad02","relation":"same-reasoning"}]},{"id":"b-1c6f61f68f50309e7d67811d","type":"p","scope":"rcnn-transfer-pipeline","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-1c6f61f68f50309e7d67811d","content":"R-CNN 先产生可能包含对象的区域，将每个区域变换为 CNN 接受的图像大小，再用 CNN 特征作区域分类与框修正。分类预训练提供的是卷积网络的参数；检测适配需要新的区域样本、前景/背景关系和位置目标。原始实现包含多个训练阶段，因此课堂画面中简洁的两项输出，不应被误读成原论文从一开始就是一个统一损失的端到端系统。","sha256":"a620c6e0f57f9d21a737c4af613fe460cac7e0d69e0cf0248c6827ec83e987db","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"R-CNN 迁移的是已学会的视觉计算","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":"rcnn-transfer-pipeline","formulas":[],"previousBlock":"b-970e13401b12f967a4320c87","nextBlock":"b-b9cd24812bfead847a6ba0a5","prerequisites":[{"id":"b-970e13401b12f967a4320c87","url":"#b-970e13401b12f967a4320c87","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"},{"id":"b-970e13401b12f967a4320c87","url":"#b-970e13401b12f967a4320c87","relation":"same-reasoning"},{"id":"b-b9cd24812bfead847a6ba0a5","url":"#b-b9cd24812bfead847a6ba0a5","relation":"same-reasoning"},{"id":"b-651424ae6acac0a4ea30ad02","url":"#b-651424ae6acac0a4ea30ad02","relation":"same-reasoning"}]},{"id":"b-b9cd24812bfead847a6ba0a5","type":"p","scope":"rcnn-transfer-pipeline","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-b9cd24812bfead847a6ba0a5","content":"为什么把区域当作一张小图很有力量？因为原来“整图里有什么”的视觉计算，现在可在许多局部窗口上复用。杯子只占整图一小部分时，候选区域让它占据输入的大部分，分类网络就能在更合适的尺度查看它。代价是裁剪、缩放与逐区域计算；区域边缘的上下文也会改变。","sha256":"861774ba5faaffdebcc9f839c46f4495cc31d8be666790b3138545d99ce00d44","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"R-CNN 迁移的是已学会的视觉计算","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":"rcnn-transfer-pipeline","formulas":[],"previousBlock":"b-1c6f61f68f50309e7d67811d","nextBlock":"b-651424ae6acac0a4ea30ad02","prerequisites":[{"id":"b-970e13401b12f967a4320c87","url":"#b-970e13401b12f967a4320c87","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"},{"id":"b-970e13401b12f967a4320c87","url":"#b-970e13401b12f967a4320c87","relation":"same-reasoning"},{"id":"b-1c6f61f68f50309e7d67811d","url":"#b-1c6f61f68f50309e7d67811d","relation":"same-reasoning"},{"id":"b-651424ae6acac0a4ea30ad02","url":"#b-651424ae6acac0a4ea30ad02","relation":"same-reasoning"}]},{"id":"b-651424ae6acac0a4ea30ad02","type":"p","scope":"rcnn-transfer-pipeline","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-651424ae6acac0a4ea30ad02","content":"原论文摘要报告 VOC 2012 检测 mAP 为 53.3%，相对之前最好结果提高超过 30%。这个同期结果支持“分类预训练与检测适配能够产生有竞争力的检测器”，不是自监督预训练的实验。作者是 Girshick、Donahue、Darrell、Malik；此处从 MoCo/MAE 接到 R-CNN，是课程按问题重组，不能反写历史先后。","sha256":"178bc2569407866ae7371a893269c1c7e57bc6d16f7f1c044f037e1730c1062d","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"R-CNN 迁移的是已学会的视觉计算","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":"rcnn-transfer-pipeline","formulas":[],"previousBlock":"b-b9cd24812bfead847a6ba0a5","nextBlock":"b-a041b1b8048c190f1d7766ed","prerequisites":[{"id":"b-970e13401b12f967a4320c87","url":"#b-970e13401b12f967a4320c87","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"},{"id":"b-970e13401b12f967a4320c87","url":"#b-970e13401b12f967a4320c87","relation":"same-reasoning"},{"id":"b-1c6f61f68f50309e7d67811d","url":"#b-1c6f61f68f50309e7d67811d","relation":"same-reasoning"},{"id":"b-b9cd24812bfead847a6ba0a5","url":"#b-b9cd24812bfead847a6ba0a5","relation":"same-reasoning"}]},{"id":"b-a041b1b8048c190f1d7766ed","type":"h3","scope":"dialogue","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-a041b1b8048c190f1d7766ed","content":"### 沿着图解追问","sha256":"f16be0424762a44344338696f06ef266c4571d15367ae2cbbd28fac98ac7e9cb","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-651424ae6acac0a4ea30ad02","nextBlock":"b-9af0333ffb0e4a95b064bad5","prerequisites":[{"id":"b-970e13401b12f967a4320c87","url":"#b-970e13401b12f967a4320c87","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]},{"id":"b-9af0333ffb0e4a95b064bad5","type":"p","scope":"dialogue","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-9af0333ffb0e4a95b064bad5","content":"以下是备课用的模拟师生对话，不是课堂实录。","sha256":"b25685f04cd1b885f3a5999e6a912460c7424021cdc5b2190584242086a82681","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-a041b1b8048c190f1d7766ed","nextBlock":"b-5039e659d4f8286848729331","prerequisites":[{"id":"b-a041b1b8048c190f1d7766ed","url":"#b-a041b1b8048c190f1d7766ed","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]},{"id":"b-5039e659d4f8286848729331","type":"p","scope":"dialogue","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-5039e659d4f8286848729331","content":"**学生：**这里为什么又用了有标签的预训练？","sha256":"f7230090f4acc46b2a1ac55d497d3030034f199ad1e9d33a7f04738dcf047b9d","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-9af0333ffb0e4a95b064bad5","nextBlock":"b-e24d120a6aed0cb7d164a0af","prerequisites":[{"id":"b-a041b1b8048c190f1d7766ed","url":"#b-a041b1b8048c190f1d7766ed","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]},{"id":"b-e24d120a6aed0cb7d164a0af","type":"p","scope":"dialogue","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-e24d120a6aed0cb7d164a0af","content":"**教师：**可迁移不是自监督专属。我们比较了学习信号，现在看已成立的监督迁移案例；逻辑连接不要求伪造历史先后。","sha256":"690ddccb15578a5c7e7b59a35026f1eef0293bd8670dccd0d6f6f600414f7eb4","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-5039e659d4f8286848729331","nextBlock":"b-4ec73e31d68009582ef971cd","prerequisites":[{"id":"b-a041b1b8048c190f1d7766ed","url":"#b-a041b1b8048c190f1d7766ed","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]},{"id":"b-4ec73e31d68009582ef971cd","type":"h2","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-4ec73e31d68009582ef971cd","content":"## 沿这个问题继续","sha256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-e24d120a6aed0cb7d164a0af","nextBlock":"b-0bc10a6781f68be89bedfc7d","prerequisites":[{"id":"b-a041b1b8048c190f1d7766ed","url":"#b-a041b1b8048c190f1d7766ed","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]},{"id":"b-0bc10a6781f68be89bedfc7d","type":"p","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-0bc10a6781f68be89bedfc7d","content":"沿 CLS07 的公式定位疑问，再选 LAB09 解析练习或 LAB10 完整训练。","sha256":"a3163e9b7618b50175088fce50e8c02bf38584dbcb466052e3f6745aa91110a4","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-4ec73e31d68009582ef971cd","nextBlock":"b-748257ec07df282a5e1777a5","prerequisites":[{"id":"b-4ec73e31d68009582ef971cd","url":"#b-4ec73e31d68009582ef971cd","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]},{"id":"b-748257ec07df282a5e1777a5","type":"h4","scope":"region-pretrained-computation","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-748257ec07df282a5e1777a5","content":"#### 图示条件与公式约定","sha256":"ef7f915c1ae63ed67bd23fecf050249273ba0efa8965d2c8d1c1c6b2d3178122","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-0bc10a6781f68be89bedfc7d","nextBlock":"b-4cd19b1fc6b80c98ae7d3dc7","prerequisites":[{"id":"b-4ec73e31d68009582ef971cd","url":"#b-4ec73e31d68009582ef971cd","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]},{"id":"b-4cd19b1fc6b80c98ae7d3dc7","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-4cd19b1fc6b80c98ae7d3dc7","content":"原始 R-CNN 使用监督分类预训练，并进行检测适配；不将教学叙述顺序误写成自监督方法的历史后继。此图省略原始多阶段训练实现。","sha256":"2e0b8ea011e1741562652d0865d8840fed4d61e7f8b1d1899b73400f0e06f2f0","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-748257ec07df282a5e1777a5","nextBlock":"b-98f1099e1f2d82e3ac7689d4","prerequisites":[{"id":"b-748257ec07df282a5e1777a5","url":"#b-748257ec07df282a5e1777a5","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]},{"id":"b-98f1099e1f2d82e3ac7689d4","type":"h4","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-98f1099e1f2d82e3ac7689d4","content":"#### 这条线索的展开与后续阅读","sha256":"48b8229f7559b36ed4d5628416e3991e5c73293fbce43374db67386c8736f3bc","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-4cd19b1fc6b80c98ae7d3dc7","nextBlock":"b-c2e93e65ad685d8a81518390","prerequisites":[{"id":"b-748257ec07df282a5e1777a5","url":"#b-748257ec07df282a5e1777a5","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]},{"id":"b-c2e93e65ad685d8a81518390","type":"ul","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-c2e93e65ad685d8a81518390","content":"- [把分类的计算，迁移到候选区域](/supplement.html?return=l3-r2-p27/0#/l3-rcnn/0) 本页说明","sha256":"b1296df119df9b111607b73a175dc1db15fc7678a5b61de3cb114f4f3711ec4c","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-98f1099e1f2d82e3ac7689d4","nextBlock":"b-97305282ac4df7da0d07b8f5","prerequisites":[{"id":"b-98f1099e1f2d82e3ac7689d4","url":"#b-98f1099e1f2d82e3ac7689d4","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]},{"id":"b-97305282ac4df7da0d07b8f5","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p27.html#b-97305282ac4df7da0d07b8f5","content":"- [把分类的计算，迁移到候选区域](/supplement.html?return=l3-r2-p27/0#/l3-rcnn/0) 本页说明","sha256":"b1296df119df9b111607b73a175dc1db15fc7678a5b61de3cb114f4f3711ec4c","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"region-computation-reuse","title":"多个区域怎样复用同一份视觉计算？"},"page":{"id":"region-pretrained-computation","title":"让预训练网络读取一个区域","number":1},"sectionTitle":"让预训练网络读取一个区域","currentQuestion":"从逐区域编码到共享整图特征，哪些计算被保留，哪些接口必须重建？","sourceSlides":["l3-r2-p27"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-c2e93e65ad685d8a81518390","nextBlock":null,"prerequisites":[{"id":"b-98f1099e1f2d82e3ac7689d4","url":"#b-98f1099e1f2d82e3ac7689d4","role":"section-title"},{"id":"b-9ee9c9a8113e4bb8ba354ff1","url":"#b-9ee9c9a8113e4bb8ba354ff1","role":"problem-context"}],"relatedBlocks":[{"id":"b-a7d8269d387b6e2a1a41998c","url":"#b-a7d8269d387b6e2a1a41998c","relation":"figure-and-explanation"}]}]}