{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/blocks/nms-measured-tradeoff.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"同一检测器改变 NMS，召回为何变好而精度变差","url":"https://codingai-lec03.pages.dev/course/blocks/nms-measured-tradeoff.html","markdown":"https://codingai-lec03.pages.dev/course/blocks/nms-measured-tradeoff.md","pages":[{"id":"reading-opening","title":"同一检测器改变 NMS，召回为何变好而精度变差","unit":"检测分数很高，错误究竟在哪里？","unitId":"detection-errors-and-evidence","chapter":"II","number":1}],"aliases":[{"oldId":"b-0f1a397ebd8f34df0474f891","targetId":"b-0f1a397ebd8f34df0474f891","page":"reading-opening","operation":"move","beforeSHA256":"ff8101cb4d8e0c925b65d60f4d3ad7e30f4a9283d59eda977d24323932d5529c","afterSHA256":"ff8101cb4d8e0c925b65d60f4d3ad7e30f4a9283d59eda977d24323932d5529c","reason":"Same source body in authored semantic page"},{"oldId":"b-906643c9bc3bde3252137a12","targetId":"b-906643c9bc3bde3252137a12","page":"reading-opening","operation":"move","beforeSHA256":"bc255c15fe7dcd27773047232605a05ecb5722a03591a1a06568309263982717","afterSHA256":"bc255c15fe7dcd27773047232605a05ecb5722a03591a1a06568309263982717","reason":"Same source body in authored semantic page"},{"oldId":"b-0f62a48f13b3ca94522804eb","targetId":"b-0f62a48f13b3ca94522804eb","page":"reading-opening","operation":"move","beforeSHA256":"8b0e46c250ed7908b144b4e40151bb579c49ca95fce5168dc35905e6cd61fbbb","afterSHA256":"8b0e46c250ed7908b144b4e40151bb579c49ca95fce5168dc35905e6cd61fbbb","reason":"Same source body in authored semantic page"},{"oldId":"b-7642b81c1dd591eee81a69d2","targetId":"b-7642b81c1dd591eee81a69d2","page":"reading-opening","operation":"rewrite","beforeSHA256":"40d8645433e6fab9cbb76b9a45aa3e1cb87f225d48a2be8bff29c6ed6d91df26","afterSHA256":"6568df1908194db0667c411a0d52480b9be36f1ec5c80e3aa45bba20e564fa18","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-504cdcc4b1fbcc96510caff8","targetId":"b-504cdcc4b1fbcc96510caff8","page":"reading-opening","operation":"move","beforeSHA256":"fa71bd1d1c49b9fd714195acb6b40c28ccf5d6455b9dd9408b7c6e7101b36997","afterSHA256":"fa71bd1d1c49b9fd714195acb6b40c28ccf5d6455b9dd9408b7c6e7101b36997","reason":"Same source body in authored semantic page"},{"oldId":"b-7603b43f747e453c4bce3d3d","targetId":"b-7603b43f747e453c4bce3d3d","page":"reading-opening","operation":"move","beforeSHA256":"07c3598d8f94057887eadde2d224c44388ad519e293cfc2ead957ce4985de933","afterSHA256":"07c3598d8f94057887eadde2d224c44388ad519e293cfc2ead957ce4985de933","reason":"Same source body in authored semantic page"},{"oldId":"b-386926290b14cb40477cada3","targetId":"b-386926290b14cb40477cada3","page":"reading-opening","operation":"move","beforeSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","afterSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","reason":"Same source body in authored semantic page"}],"blocks":[{"id":"b-0f1a397ebd8f34df0474f891","type":"h1","scope":"nms-measured-tradeoff","url":"https://codingai-lec03.pages.dev/course/blocks/nms-measured-tradeoff.html#b-0f1a397ebd8f34df0474f891","content":"# 同一检测器改变 NMS，召回为何变好而精度变差","sha256":"ff8101cb4d8e0c925b65d60f4d3ad7e30f4a9283d59eda977d24323932d5529c","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"reading-opening","title":"同一检测器改变 NMS，召回为何变好而精度变差","number":1},"sectionTitle":"同一检测器改变 NMS，召回为何变好而精度变差","currentQuestion":"同一检测器改变 NMS，召回为何变好而精度变差","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-906643c9bc3bde3252137a12","prerequisites":[{"id":"b-906643c9bc3bde3252137a12","url":"#b-906643c9bc3bde3252137a12","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-906643c9bc3bde3252137a12","type":"p","scope":"nms-measured-tradeoff","url":"https://codingai-lec03.pages.dev/course/blocks/nms-measured-tradeoff.html#b-906643c9bc3bde3252137a12","content":"三个重叠预测可能对应一只杯子。NMS 用预测之间的分数与重叠整理重复，评价匹配则使用真值，二者不是同一操作。先看三框算例，再看同一真实教学检测器放宽 NMS 后召回和精度的相反变化。","sha256":"bc255c15fe7dcd27773047232605a05ecb5722a03591a1a06568309263982717","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"reading-opening","title":"同一检测器改变 NMS，召回为何变好而精度变差","number":1},"sectionTitle":"同一检测器改变 NMS，召回为何变好而精度变差","currentQuestion":"同一检测器改变 NMS，召回为何变好而精度变差","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-0f1a397ebd8f34df0474f891","nextBlock":"b-0f62a48f13b3ca94522804eb","prerequisites":[{"id":"b-0f1a397ebd8f34df0474f891","url":"#b-0f1a397ebd8f34df0474f891","role":"section-title"}],"relatedBlocks":[]},{"id":"b-0f62a48f13b3ca94522804eb","type":"p","scope":"nms-measured-tradeoff","url":"https://codingai-lec03.pages.dev/course/blocks/nms-measured-tradeoff.html#b-0f62a48f13b3ca94522804eb","content":"LAB10 已保存一组更接近真实使用的对照：输入与训练后权重保持不变，只把 NMS IoU 阈值从 .35 放宽到 .65。展示/工作点的分数门槛固定 .4。","sha256":"8b0e46c250ed7908b144b4e40151bb579c49ca95fce5168dc35905e6cd61fbbb","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"reading-opening","title":"同一检测器改变 NMS，召回为何变好而精度变差","number":1},"sectionTitle":"同一检测器改变 NMS，召回为何变好而精度变差","currentQuestion":"同一检测器改变 NMS，召回为何变好而精度变差","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-906643c9bc3bde3252137a12","nextBlock":"b-7642b81c1dd591eee81a69d2","prerequisites":[{"id":"b-0f1a397ebd8f34df0474f891","url":"#b-0f1a397ebd8f34df0474f891","role":"section-title"},{"id":"b-906643c9bc3bde3252137a12","url":"#b-906643c9bc3bde3252137a12","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7642b81c1dd591eee81a69d2","type":"table","scope":"nms-measured-tradeoff","url":"https://codingai-lec03.pages.dev/course/blocks/nms-measured-tradeoff.html#b-7642b81c1dd591eee81a69d2","content":"| NMS IoU | recall@工作点 | precision@工作点 | AP50 |\n| --- | --- | --- | --- |\n| .35 | .98311 | .94175 | .97303 |\n| .65 | 1.00000 | .41926 | .95866 |","sha256":"6568df1908194db0667c411a0d52480b9be36f1ec5c80e3aa45bba20e564fa18","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"reading-opening","title":"同一检测器改变 NMS，召回为何变好而精度变差","number":1},"sectionTitle":"同一检测器改变 NMS，召回为何变好而精度变差","currentQuestion":"同一检测器改变 NMS，召回为何变好而精度变差","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"table":{"headers":["NMS IoU","recall@工作点","precision@工作点","AP50"],"rows":[["NMS IoU","recall@工作点","precision@工作点","AP50"],[".35",".98311",".94175",".97303"],[".65","1.00000",".41926",".95866"]]},"previousBlock":"b-0f62a48f13b3ca94522804eb","nextBlock":"b-504cdcc4b1fbcc96510caff8","prerequisites":[{"id":"b-0f1a397ebd8f34df0474f891","url":"#b-0f1a397ebd8f34df0474f891","role":"section-title"},{"id":"b-906643c9bc3bde3252137a12","url":"#b-906643c9bc3bde3252137a12","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-504cdcc4b1fbcc96510caff8","type":"p","scope":"nms-measured-tradeoff","url":"https://codingai-lec03.pages.dev/course/blocks/nms-measured-tradeoff.html#b-504cdcc4b1fbcc96510caff8","content":"放宽后，一些原来被抑制的候选重新出现，补到了最后少数对象；同时大量同一对象的重复框留在输出，所以 precision 大降。AP 的变化比这个固定工作点的 precision 小，因为 AP 汇总整个分数排序，而工作点只看分数至少 .4 的输出。","sha256":"fa71bd1d1c49b9fd714195acb6b40c28ccf5d6455b9dd9408b7c6e7101b36997","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"reading-opening","title":"同一检测器改变 NMS，召回为何变好而精度变差","number":1},"sectionTitle":"同一检测器改变 NMS，召回为何变好而精度变差","currentQuestion":"同一检测器改变 NMS，召回为何变好而精度变差","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7642b81c1dd591eee81a69d2","nextBlock":"b-7603b43f747e453c4bce3d3d","prerequisites":[{"id":"b-0f1a397ebd8f34df0474f891","url":"#b-0f1a397ebd8f34df0474f891","role":"section-title"},{"id":"b-906643c9bc3bde3252137a12","url":"#b-906643c9bc3bde3252137a12","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7603b43f747e453c4bce3d3d","type":"p","scope":"nms-measured-tradeoff","url":"https://codingai-lec03.pages.dev/course/blocks/nms-measured-tradeoff.html#b-7603b43f747e453c4bce3d3d","content":"这组已有本地结果不是本轮新运行。它的价值是把 P34 的两种效应落实到同一模型：预测数量增加本身没有确定好坏，必须看新增的是哪个对象、是否重复，以及部署时怎样使用阈值。选阈值应使用验证集与任务成本；不能反复挑保留集最漂亮的数字再称泛化结果。","sha256":"07c3598d8f94057887eadde2d224c44388ad519e293cfc2ead957ce4985de933","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"reading-opening","title":"同一检测器改变 NMS，召回为何变好而精度变差","number":1},"sectionTitle":"同一检测器改变 NMS，召回为何变好而精度变差","currentQuestion":"同一检测器改变 NMS，召回为何变好而精度变差","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-504cdcc4b1fbcc96510caff8","nextBlock":"b-386926290b14cb40477cada3","prerequisites":[{"id":"b-0f1a397ebd8f34df0474f891","url":"#b-0f1a397ebd8f34df0474f891","role":"section-title"},{"id":"b-906643c9bc3bde3252137a12","url":"#b-906643c9bc3bde3252137a12","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-386926290b14cb40477cada3","type":"h2","scope":"nms-measured-tradeoff","url":"https://codingai-lec03.pages.dev/course/blocks/nms-measured-tradeoff.html#b-386926290b14cb40477cada3","content":"## 继续阅读与实验","sha256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"detection-errors-and-evidence","title":"检测分数很高，错误究竟在哪里？"},"page":{"id":"reading-opening","title":"同一检测器改变 NMS，召回为何变好而精度变差","number":1},"sectionTitle":"同一检测器改变 NMS，召回为何变好而精度变差","currentQuestion":"同一检测器改变 NMS，召回为何变好而精度变差","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7603b43f747e453c4bce3d3d","nextBlock":null,"prerequisites":[{"id":"b-0f1a397ebd8f34df0474f891","url":"#b-0f1a397ebd8f34df0474f891","role":"section-title"},{"id":"b-906643c9bc3bde3252137a12","url":"#b-906643c9bc3bde3252137a12","role":"problem-context"}],"relatedBlocks":[]}]}