{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/blocks/nms-versus-matching.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"NMS 不看真值，评价匹配必须看真值","url":"https://codingai-lec03.pages.dev/course/blocks/nms-versus-matching.html","markdown":"https://codingai-lec03.pages.dev/course/blocks/nms-versus-matching.md","pages":[{"id":"reading-opening","title":"NMS 不看真值，评价匹配必须看真值","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-a2d83a353c9e03064690685a","targetId":"b-a2d83a353c9e03064690685a","page":"reading-opening","operation":"move","beforeSHA256":"f0849b96d215bc21d5f7a315db8ce00d67e13a1a53d3f00f32a216723ccdef98","afterSHA256":"f0849b96d215bc21d5f7a315db8ce00d67e13a1a53d3f00f32a216723ccdef98","reason":"Same source body in authored semantic page"},{"oldId":"b-8497ea764c18348251aa2d7e","targetId":"b-8497ea764c18348251aa2d7e","page":"reading-opening","operation":"move","beforeSHA256":"bc255c15fe7dcd27773047232605a05ecb5722a03591a1a06568309263982717","afterSHA256":"bc255c15fe7dcd27773047232605a05ecb5722a03591a1a06568309263982717","reason":"Same source body in authored semantic page"},{"oldId":"b-f5593872e2e22e47ab7de9aa","targetId":"b-f5593872e2e22e47ab7de9aa","page":"reading-opening","operation":"move","beforeSHA256":"69c7a61af806c6ca9de4345591f7a05bbcbf8db6b0d05d12812921088825707b","afterSHA256":"69c7a61af806c6ca9de4345591f7a05bbcbf8db6b0d05d12812921088825707b","reason":"Same source body in authored semantic page"},{"oldId":"b-aa29d224e88f306659a6e8b1","targetId":"b-aa29d224e88f306659a6e8b1","page":"reading-opening","operation":"move","beforeSHA256":"b9835ac8c66d97cf6a7f6cfc2ecf3b926d4b1d59a818684234e480ec9282c233","afterSHA256":"b9835ac8c66d97cf6a7f6cfc2ecf3b926d4b1d59a818684234e480ec9282c233","reason":"Same source body in authored semantic page"},{"oldId":"b-aefc71bb690386e4156ea0cd","targetId":"b-aefc71bb690386e4156ea0cd","page":"reading-opening","operation":"move","beforeSHA256":"be1d7ca5463e609cb8b7bca3d6578ee106e7dc449c001fac0f565b86e830e8ef","afterSHA256":"be1d7ca5463e609cb8b7bca3d6578ee106e7dc449c001fac0f565b86e830e8ef","reason":"Same source body in authored semantic page"},{"oldId":"b-3f90459669330949e715c7bd","targetId":"b-3f90459669330949e715c7bd","page":"reading-opening","operation":"move","beforeSHA256":"fc327f3830cedf2a67ebb5483684e49c5db347ab528fa495a762bdac99622e32","afterSHA256":"fc327f3830cedf2a67ebb5483684e49c5db347ab528fa495a762bdac99622e32","reason":"Same source body in authored semantic page"},{"oldId":"b-74fd1311940a3490a7122487","targetId":"b-74fd1311940a3490a7122487","page":"reading-opening","operation":"move","beforeSHA256":"ebd9b9fec993f0371d99690a74d56ce7c4f0cbcb5184361685a3e4a87472bef2","afterSHA256":"ebd9b9fec993f0371d99690a74d56ce7c4f0cbcb5184361685a3e4a87472bef2","reason":"Same source body in authored semantic page"},{"oldId":"b-b20353c2cb4968fbefb93901","targetId":"b-b20353c2cb4968fbefb93901","page":"reading-opening","operation":"rewrite","beforeSHA256":"cae41d1c2c24c9cd1e14a08f789d55a7e56b6da8aa0c76dfd9ad11ac9582b0e9","afterSHA256":"41c41fc14c453eef368036a149116f446cb24977f4c00b6ffc018b655f68b191","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-1ff7be77657cc03830c6b712","targetId":"b-1ff7be77657cc03830c6b712","page":"reading-opening","operation":"move","beforeSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","afterSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","reason":"Same source body in authored semantic page"}],"blocks":[{"id":"b-a2d83a353c9e03064690685a","type":"h1","scope":"nms-versus-matching","url":"https://codingai-lec03.pages.dev/course/blocks/nms-versus-matching.html#b-a2d83a353c9e03064690685a","content":"# NMS 不看真值，评价匹配必须看真值","sha256":"f0849b96d215bc21d5f7a315db8ce00d67e13a1a53d3f00f32a216723ccdef98","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":"NMS 不看真值，评价匹配必须看真值","number":1},"sectionTitle":"NMS 不看真值，评价匹配必须看真值","currentQuestion":"NMS 不看真值，评价匹配必须看真值","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-8497ea764c18348251aa2d7e","prerequisites":[{"id":"b-8497ea764c18348251aa2d7e","url":"#b-8497ea764c18348251aa2d7e","role":"problem-context"}],"relatedBlocks":[{"id":"b-b20353c2cb4968fbefb93901","url":"#b-b20353c2cb4968fbefb93901","relation":"figure-and-explanation"}]},{"id":"b-8497ea764c18348251aa2d7e","type":"p","scope":"nms-versus-matching","url":"https://codingai-lec03.pages.dev/course/blocks/nms-versus-matching.html#b-8497ea764c18348251aa2d7e","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":"NMS 不看真值，评价匹配必须看真值","number":1},"sectionTitle":"NMS 不看真值，评价匹配必须看真值","currentQuestion":"NMS 不看真值，评价匹配必须看真值","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-a2d83a353c9e03064690685a","nextBlock":"b-f5593872e2e22e47ab7de9aa","prerequisites":[{"id":"b-a2d83a353c9e03064690685a","url":"#b-a2d83a353c9e03064690685a","role":"section-title"}],"relatedBlocks":[{"id":"b-b20353c2cb4968fbefb93901","url":"#b-b20353c2cb4968fbefb93901","relation":"figure-and-explanation"}]},{"id":"b-f5593872e2e22e47ab7de9aa","type":"p","scope":"nms-versus-matching","url":"https://codingai-lec03.pages.dev/course/blocks/nms-versus-matching.html#b-f5593872e2e22e47ab7de9aa","content":"NMS 从预测框中整理重复。按类别与分数排序，保留最高分框 A，再抑制与 A 的 IoU 大于阈值的同类低分框，然后继续。输入只有预测框、分数和类别，不需要人工真值。","sha256":"69c7a61af806c6ca9de4345591f7a05bbcbf8db6b0d05d12812921088825707b","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":"NMS 不看真值，评价匹配必须看真值","number":1},"sectionTitle":"NMS 不看真值，评价匹配必须看真值","currentQuestion":"NMS 不看真值，评价匹配必须看真值","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-8497ea764c18348251aa2d7e","nextBlock":"b-aa29d224e88f306659a6e8b1","prerequisites":[{"id":"b-a2d83a353c9e03064690685a","url":"#b-a2d83a353c9e03064690685a","role":"section-title"},{"id":"b-8497ea764c18348251aa2d7e","url":"#b-8497ea764c18348251aa2d7e","role":"problem-context"}],"relatedBlocks":[{"id":"b-b20353c2cb4968fbefb93901","url":"#b-b20353c2cb4968fbefb93901","relation":"figure-and-explanation"}]},{"id":"b-aa29d224e88f306659a6e8b1","type":"p","scope":"nms-versus-matching","url":"https://codingai-lec03.pages.dev/course/blocks/nms-versus-matching.html#b-aa29d224e88f306659a6e8b1","content":"回到 P34 同一杯子上的三个框。A、B、C 的教学分数分别为 0.90、0.80、0.60；IoU(A,B)≈0.90，IoU(A,C)≈0.88。按分数先保留 A，在 NMS 阈值为 0.5 时，B 和 C 都与 A 过度重叠，因此都被抑制。蓝杯的另一个框与 A 不重叠，仍然保留。这里比较的是预测框彼此之间的重叠，还没有使用真值。","sha256":"b9835ac8c66d97cf6a7f6cfc2ecf3b926d4b1d59a818684234e480ec9282c233","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":"NMS 不看真值，评价匹配必须看真值","number":1},"sectionTitle":"NMS 不看真值，评价匹配必须看真值","currentQuestion":"NMS 不看真值，评价匹配必须看真值","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f5593872e2e22e47ab7de9aa","nextBlock":"b-aefc71bb690386e4156ea0cd","prerequisites":[{"id":"b-a2d83a353c9e03064690685a","url":"#b-a2d83a353c9e03064690685a","role":"section-title"},{"id":"b-8497ea764c18348251aa2d7e","url":"#b-8497ea764c18348251aa2d7e","role":"problem-context"}],"relatedBlocks":[{"id":"b-b20353c2cb4968fbefb93901","url":"#b-b20353c2cb4968fbefb93901","relation":"figure-and-explanation"}]},{"id":"b-aefc71bb690386e4156ea0cd","type":"p","scope":"nms-versus-matching","url":"https://codingai-lec03.pages.dev/course/blocks/nms-versus-matching.html#b-aefc71bb690386e4156ea0cd","content":"若把阈值放宽到 0.90，在本例原始坐标下，IoU(A,B)=0.8950778、IoU(A,C)=0.8774870，都没有超过阈值；重复框便可能继续留下。因此判定应使用未四舍五入的 IoU，图中 0.90 只是显示精度。下一步评价才用真值判断它们是在补漏检还是重复命中。","sha256":"be1d7ca5463e609cb8b7bca3d6578ee106e7dc449c001fac0f565b86e830e8ef","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":"NMS 不看真值，评价匹配必须看真值","number":1},"sectionTitle":"NMS 不看真值，评价匹配必须看真值","currentQuestion":"NMS 不看真值，评价匹配必须看真值","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-aa29d224e88f306659a6e8b1","nextBlock":"b-3f90459669330949e715c7bd","prerequisites":[{"id":"b-a2d83a353c9e03064690685a","url":"#b-a2d83a353c9e03064690685a","role":"section-title"},{"id":"b-8497ea764c18348251aa2d7e","url":"#b-8497ea764c18348251aa2d7e","role":"problem-context"}],"relatedBlocks":[{"id":"b-b20353c2cb4968fbefb93901","url":"#b-b20353c2cb4968fbefb93901","relation":"figure-and-explanation"}]},{"id":"b-3f90459669330949e715c7bd","type":"p","scope":"nms-versus-matching","url":"https://codingai-lec03.pages.dev/course/blocks/nms-versus-matching.html#b-3f90459669330949e715c7bd","content":"若两只真实杯子互相遮挡，两个正确框也可能高度重叠。将 `(0,0,2,2)` 与 `(.5,0,2.5,2)` 当作两只对象的框，交集 3、并集 5，IoU=.6；NMS=.5 就可能删掉较低分的真对象。它以几何重叠近似重复，没有直接知道对象身份。","sha256":"fc327f3830cedf2a67ebb5483684e49c5db347ab528fa495a762bdac99622e32","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":"NMS 不看真值，评价匹配必须看真值","number":1},"sectionTitle":"NMS 不看真值，评价匹配必须看真值","currentQuestion":"NMS 不看真值，评价匹配必须看真值","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-aefc71bb690386e4156ea0cd","nextBlock":"b-74fd1311940a3490a7122487","prerequisites":[{"id":"b-a2d83a353c9e03064690685a","url":"#b-a2d83a353c9e03064690685a","role":"section-title"},{"id":"b-8497ea764c18348251aa2d7e","url":"#b-8497ea764c18348251aa2d7e","role":"problem-context"}],"relatedBlocks":[{"id":"b-b20353c2cb4968fbefb93901","url":"#b-b20353c2cb4968fbefb93901","relation":"figure-and-explanation"}]},{"id":"b-74fd1311940a3490a7122487","type":"p","scope":"nms-versus-matching","url":"https://codingai-lec03.pages.dev/course/blocks/nms-versus-matching.html#b-74fd1311940a3490a7122487","content":"评价再用保留下来的框与人工对象对应。重复命中同一真值只允许一次计为 TP，其余是 FP；漏掉的对象降低 recall。模型预测、后处理、评价三层分开，就能解释为什么“放宽 NMS 得到更多框”既可能补召回，也可能伤精度。","sha256":"ebd9b9fec993f0371d99690a74d56ce7c4f0cbcb5184361685a3e4a87472bef2","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":"NMS 不看真值，评价匹配必须看真值","number":1},"sectionTitle":"NMS 不看真值，评价匹配必须看真值","currentQuestion":"NMS 不看真值，评价匹配必须看真值","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-3f90459669330949e715c7bd","nextBlock":"b-b20353c2cb4968fbefb93901","prerequisites":[{"id":"b-a2d83a353c9e03064690685a","url":"#b-a2d83a353c9e03064690685a","role":"section-title"},{"id":"b-8497ea764c18348251aa2d7e","url":"#b-8497ea764c18348251aa2d7e","role":"problem-context"}],"relatedBlocks":[{"id":"b-b20353c2cb4968fbefb93901","url":"#b-b20353c2cb4968fbefb93901","relation":"figure-and-explanation"}]},{"id":"b-b20353c2cb4968fbefb93901","type":"figure","scope":"nms-versus-matching","url":"https://codingai-lec03.pages.dev/course/blocks/nms-versus-matching.html#b-b20353c2cb4968fbefb93901","content":"图内标签与数值：Lecture 03 课堂图解 P34；保留 A；B、C 与它过度重叠；教学分数 / IoU 阈值 0.5；A  0.90   保留；B  0.80   抑制；C  0.60   抑制；另一只杯子仍然保留；拥挤对象也可能互相压制；按同类别分数排序，抑制 IoU > 0.5 的重叠预测；分数是教学设定。\n\n同一红杯上的 A、B、C：阈值0.5保留A、抑制B和C；蓝杯独立框继续保留。比较预测之间的IoU，不使用真值。 查看对应图解状态\n\n- 保留 A；B、C 与它过度重叠\n- 教学分数 / IoU 阈值 0.5\n- A 0.90 保留\n- B 0.80 抑制\n- C 0.60 抑制\n- 另一只杯子仍然保留\n- 拥挤对象也可能互相压制\n- 按同类别分数排序，抑制 IoU > 0.5 的重叠预测；分数是教学设定。同一红杯上的 A、B、C：阈值0.5保留A、抑制B和C；蓝杯独立框继续保留。比较预测之间的IoU，不使用真值。 [查看对应图解状态](/#/l3-r2-p34/1)[查看完整图解](/#/l3-r2-p34/1)","sha256":"41c41fc14c453eef368036a149116f446cb24977f4c00b6ffc018b655f68b191","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":"NMS 不看真值，评价匹配必须看真值","number":1},"sectionTitle":"NMS 不看真值，评价匹配必须看真值","currentQuestion":"NMS 不看真值，评价匹配必须看真值","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"同一红杯上的 A、B、C：阈值0.5保留A、抑制B和C；蓝杯独立框继续保留。比较预测之间的IoU，不使用真值。 查看对应图解状态","labels":["保留 A；B、C 与它过度重叠","教学分数 / IoU 阈值 0.5","A  0.90   保留","B  0.80   抑制","C  0.60   抑制","另一只杯子仍然保留","拥挤对象也可能互相压制","按同类别分数排序，抑制 IoU > 0.5 的重叠预测；分数是教学设定。"],"description":"同一红杯上的 A、B、C：阈值0.5保留A、抑制B和C；蓝杯独立框继续保留。比较预测之间的IoU，不使用真值。 查看对应图解状态保留 A；B、C 与它过度重叠教学分数 / IoU 阈值 0.5A 0.90 保留B 0.80 抑制C 0.60 抑制另一只杯子仍然保留拥挤对象也可能互相压制按同类别分数排序，抑制 IoU > 0.5 的重叠预测；分数是教学设定。","evidenceIdentity":"同一红杯上的 A、B、C：阈值0.5保留A、抑制B和C；蓝杯独立框继续保留。比较预测之间的IoU，不使用真值。 查看对应图解状态","images":[{"src":"/assets/display/8ced939a1fd55047fbf73ea5.webp","alt":""}]},"previousBlock":"b-74fd1311940a3490a7122487","nextBlock":"b-1ff7be77657cc03830c6b712","prerequisites":[{"id":"b-a2d83a353c9e03064690685a","url":"#b-a2d83a353c9e03064690685a","role":"section-title"},{"id":"b-8497ea764c18348251aa2d7e","url":"#b-8497ea764c18348251aa2d7e","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-1ff7be77657cc03830c6b712","type":"h2","scope":"nms-versus-matching","url":"https://codingai-lec03.pages.dev/course/blocks/nms-versus-matching.html#b-1ff7be77657cc03830c6b712","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":"NMS 不看真值，评价匹配必须看真值","number":1},"sectionTitle":"NMS 不看真值，评价匹配必须看真值","currentQuestion":"NMS 不看真值，评价匹配必须看真值","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b20353c2cb4968fbefb93901","nextBlock":null,"prerequisites":[{"id":"b-a2d83a353c9e03064690685a","url":"#b-a2d83a353c9e03064690685a","role":"section-title"},{"id":"b-8497ea764c18348251aa2d7e","url":"#b-8497ea764c18348251aa2d7e","role":"problem-context"}],"relatedBlocks":[{"id":"b-b20353c2cb4968fbefb93901","url":"#b-b20353c2cb4968fbefb93901","relation":"figure-and-explanation"}]}]}