{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"NMS 与评价匹配解决两个不同的重复问题","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html","markdown":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.md","pages":[{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","unit":"NMS 与评价匹配解决两个不同的重复问题","unitId":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","chapter":"course","number":1},{"id":"reading-section-1","title":"回看：共享卷积、RPN 与一阶段检测的计算位置","unit":"NMS 与评价匹配解决两个不同的重复问题","unitId":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","chapter":"course","number":2}],"aliases":[{"oldId":"b-f0723f2ffab9d45507f8a7fe","targetId":"b-f0723f2ffab9d45507f8a7fe","operation":"historical-explanation","sourceScope":"document"},{"oldId":"b-26823644ad1469850d4fd955","targetId":"b-26823644ad1469850d4fd955","operation":"historical-explanation","sourceScope":"document"},{"oldId":"b-e039ebc0d0ea92637529691d","targetId":"b-e039ebc0d0ea92637529691d","operation":"historical-explanation","sourceScope":"document"},{"oldId":"b-274406b52c91ccc80bcd2f78","targetId":"b-274406b52c91ccc80bcd2f78","operation":"historical-explanation","sourceScope":"document"},{"oldId":"b-58dd2fe43e3d2dd60230b0cc","targetId":"b-58dd2fe43e3d2dd60230b0cc","page":"reading-opening","operation":"move","beforeSHA256":"dd3b43818a82a1bb184e937a82a3c3ddecae57c2f7d0f76651320e47423d8548","afterSHA256":"dd3b43818a82a1bb184e937a82a3c3ddecae57c2f7d0f76651320e47423d8548","reason":"Same source body in authored semantic page"},{"oldId":"b-5e59d705a75c738ab1a89dd6","targetId":"b-5e59d705a75c738ab1a89dd6","page":"reading-opening","operation":"move","beforeSHA256":"ecda282b239c3074e300248265fdf554e98dd5cfa46ae2b72033e389236ddeb4","afterSHA256":"ecda282b239c3074e300248265fdf554e98dd5cfa46ae2b72033e389236ddeb4","reason":"Same source body in authored semantic page"},{"oldId":"b-1b722117caa5e008be58d3fb","targetId":"b-1b722117caa5e008be58d3fb","page":"reading-opening","operation":"move","beforeSHA256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","afterSHA256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","reason":"Same source body in authored semantic page"},{"oldId":"b-9fc9426c423d234ee83ee5c8","targetId":"b-9fc9426c423d234ee83ee5c8","page":"reading-opening","operation":"move","beforeSHA256":"88216e858779670a15bd42f20a06f136275ceee6a8685216f5821af4b1faf48f","afterSHA256":"88216e858779670a15bd42f20a06f136275ceee6a8685216f5821af4b1faf48f","reason":"Same source body in authored semantic page"},{"oldId":"b-d5541ce223b78d216f754a6a","targetId":"b-d5541ce223b78d216f754a6a","page":"reading-opening","operation":"move","beforeSHA256":"72d3c7834c3fafe18571e3ac3fc55ec05967604ca8e1c604fc52313b1c6b033b","afterSHA256":"72d3c7834c3fafe18571e3ac3fc55ec05967604ca8e1c604fc52313b1c6b033b","reason":"Same source body in authored semantic page"},{"oldId":"b-4cd607b6283ba152137aee5c","targetId":"b-4cd607b6283ba152137aee5c","page":"reading-opening","operation":"move","beforeSHA256":"b117cf446d2ad20730ef6f594c83df176d7660e6471a02cd3f1b71e3e15edf9e","afterSHA256":"b117cf446d2ad20730ef6f594c83df176d7660e6471a02cd3f1b71e3e15edf9e","reason":"Same source body in authored semantic page"},{"oldId":"b-d92e5777909882aadc531f81","targetId":"b-d92e5777909882aadc531f81","page":"reading-opening","operation":"move","beforeSHA256":"315a10d85c66d1a4cc67a41fd3d34c08e2b14631a03e3d0174dc9b1de08132f4","afterSHA256":"315a10d85c66d1a4cc67a41fd3d34c08e2b14631a03e3d0174dc9b1de08132f4","reason":"Same source body in authored semantic page"},{"oldId":"b-afc4aee1cb13579c45453fb3","targetId":"b-afc4aee1cb13579c45453fb3","page":"reading-opening","operation":"move","beforeSHA256":"ac5d224648f4ebde30f06c9a67e260c09658b54cddb2f0ff694279ac674969bc","afterSHA256":"ac5d224648f4ebde30f06c9a67e260c09658b54cddb2f0ff694279ac674969bc","reason":"Same source body in authored semantic page"},{"oldId":"b-7e7918882942596f8b268e6b","targetId":"b-7e7918882942596f8b268e6b","page":"reading-opening","operation":"move","beforeSHA256":"fdf90e10bf1db95f074762e097e80c61cd82e960b2ac2808385f306853065c3c","afterSHA256":"fdf90e10bf1db95f074762e097e80c61cd82e960b2ac2808385f306853065c3c","reason":"Same source body in authored semantic page"},{"oldId":"b-467c3a2fcc3349fc3731f705","targetId":"b-467c3a2fcc3349fc3731f705","page":"reading-opening","operation":"move","beforeSHA256":"9f8de4572c912ab854fcccf5932a372018412e10c44dbb237cf7260d10cfc308","afterSHA256":"9f8de4572c912ab854fcccf5932a372018412e10c44dbb237cf7260d10cfc308","reason":"Same source body in authored semantic page"},{"oldId":"b-f7d2cbee72c87aec06cd8b22","targetId":"b-f7d2cbee72c87aec06cd8b22","page":"reading-opening","operation":"move","beforeSHA256":"79ae2f09f6174dcb8948392160784f39aefcad99854bef12aef60b9ed330c357","afterSHA256":"79ae2f09f6174dcb8948392160784f39aefcad99854bef12aef60b9ed330c357","reason":"Same source body in authored semantic page"},{"oldId":"b-5b5fbb7ac469d758aeb27b88","targetId":"b-5b5fbb7ac469d758aeb27b88","page":"reading-opening","operation":"move","beforeSHA256":"939c4761628c5da65993b40a6d844bfb7378ef29dbeeb37b428d08c76b31f013","afterSHA256":"939c4761628c5da65993b40a6d844bfb7378ef29dbeeb37b428d08c76b31f013","reason":"Same source body in authored semantic page"},{"oldId":"b-dd5228970e460ce1a5b60017","targetId":"b-dd5228970e460ce1a5b60017","page":"reading-opening","operation":"rewrite","beforeSHA256":"5c0fc9438aff52b91880680acb5caaf5089e0667dc765caf1933d42591dad233","afterSHA256":"f24808c9a952fb28fdcdc02be7ca78258ce96c4f5df42045c6c1f671959e3cb0","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-67205fa5b0862ac5f951d0d3","targetId":"b-67205fa5b0862ac5f951d0d3","page":"reading-opening","operation":"move","beforeSHA256":"8a25829866cb20b99f921364d86c33657dae9e1547549d9560d15ceb1a6f9537","afterSHA256":"8a25829866cb20b99f921364d86c33657dae9e1547549d9560d15ceb1a6f9537","reason":"Same source body in authored semantic page"},{"oldId":"b-9864a3ae81d5531e1d91a1d1","targetId":"b-9864a3ae81d5531e1d91a1d1","page":"reading-opening","operation":"move","beforeSHA256":"9d2d37fbdcea1339fbcd1aeaef278a3912086f9ddca5746afa81e9059b4bd22b","afterSHA256":"09e76a6115a60db2f769dcaf9677a8e9e048c037ec149a8c650a05d73a12af9f","reason":"Same source body in authored semantic page"},{"oldId":"b-0857a6d2d9beea2f1cdc55ad","targetId":"b-0857a6d2d9beea2f1cdc55ad","page":"reading-opening","operation":"move","beforeSHA256":"f0c22e62c5520d432ca61fc6bad6c1e01e0460e70f7f9f2bdc6150386fd7aa62","afterSHA256":"9ab6840d67cc4c3ac9ff62f6ec9a8c36064f02605214b40349dbf9191fcbd589","reason":"Same source body in authored semantic page"},{"oldId":"b-19733dca83a79fd874714af1","targetId":"b-19733dca83a79fd874714af1","page":"reading-opening","operation":"move","beforeSHA256":"d50b94eb722dee5ced832487e587531ad8fc0c7e1ea6a13ff5fe4d0906e93d7b","afterSHA256":"d50b94eb722dee5ced832487e587531ad8fc0c7e1ea6a13ff5fe4d0906e93d7b","reason":"Same source body in authored semantic page"},{"oldId":"b-42389b57dbaa0d6b5d24f7b4","targetId":"b-42389b57dbaa0d6b5d24f7b4","page":"reading-opening","operation":"move","beforeSHA256":"8b7d63701e0e921bee18b7f1c1e38d6db5d9a5db3cb2e40577eaa14a66c80021","afterSHA256":"8b7d63701e0e921bee18b7f1c1e38d6db5d9a5db3cb2e40577eaa14a66c80021","reason":"Same source body in authored semantic page"},{"oldId":"b-725960febe596dd5ec241fd8","targetId":"b-725960febe596dd5ec241fd8","page":"reading-opening","operation":"move","beforeSHA256":"47399423a3cd722f6cc81989106117ca9567f5f15f0b56d84fbad77bbc4db80c","afterSHA256":"47399423a3cd722f6cc81989106117ca9567f5f15f0b56d84fbad77bbc4db80c","reason":"Same source body in authored semantic page"},{"oldId":"b-f8fc7263b378ee3f981acb51","targetId":"b-f8fc7263b378ee3f981acb51","page":"reading-section-1","operation":"move","beforeSHA256":"11cc60ecac5230cad0208d185147fdee5e90412fc3a83335039e6be3103df597","afterSHA256":"11cc60ecac5230cad0208d185147fdee5e90412fc3a83335039e6be3103df597","reason":"Same source body in authored semantic page"},{"oldId":"b-505e617fd7c6420f7c20ecfd","targetId":"b-505e617fd7c6420f7c20ecfd","page":"reading-section-1","operation":"move","beforeSHA256":"bee32e87d504e0167d898e9c0f91a56e4c7b091c9b8feb49d487de4dab842691","afterSHA256":"bee32e87d504e0167d898e9c0f91a56e4c7b091c9b8feb49d487de4dab842691","reason":"Same source body in authored semantic page"},{"oldId":"b-f42c34c02f11c1ec409055b0","targetId":"b-f42c34c02f11c1ec409055b0","page":"reading-section-1","operation":"move","beforeSHA256":"4746f9c7084f10c1639a38d91b9ad149f85aba184e86b9c42f324db145121c64","afterSHA256":"4746f9c7084f10c1639a38d91b9ad149f85aba184e86b9c42f324db145121c64","reason":"Same source body in authored semantic page"},{"oldId":"b-028f808e19496034f92c0528","targetId":"b-028f808e19496034f92c0528","page":"reading-section-1","operation":"move","beforeSHA256":"609abc3860560a383fa699b0b71970390c0f3d0f80d59952beb64f4d04b0862a","afterSHA256":"609abc3860560a383fa699b0b71970390c0f3d0f80d59952beb64f4d04b0862a","reason":"Same source body in authored semantic page"},{"oldId":"block-8f46ef44ed4e4529beedf33817f14d3f--nms-与评价匹配解决两个不同的重复问题","targetId":"b-d5541ce223b78d216f754a6a","operation":"retain-source-identity"},{"oldId":"block-8f46ef44ed4e4529beedf33817f14d3f--沿一次真实训练追踪从图像到框","targetId":"b-467c3a2fcc3349fc3731f705","operation":"retain-source-identity"},{"oldId":"block-8f46ef44ed4e4529beedf33817f14d3f--共享特征减少重复计算但对象身份仍要保留","targetId":"b-f8fc7263b378ee3f981acb51","operation":"retain-source-identity"}],"blocks":[{"id":"b-58dd2fe43e3d2dd60230b0cc","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-58dd2fe43e3d2dd60230b0cc","content":"[CLS 07 · 检测几何、匹配与评价](/course/notebooks/586e66e049764158a50747e69c219312.html)","sha256":"dd3b43818a82a1bb184e937a82a3c3ddecae57c2f7d0f76651320e47423d8548","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-5e59d705a75c738ab1a89dd6","prerequisites":[],"relatedBlocks":[]},{"id":"b-5e59d705a75c738ab1a89dd6","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-5e59d705a75c738ab1a89dd6","content":"# NMS 与评价匹配解决两个不同的重复问题","sha256":"ecda282b239c3074e300248265fdf554e98dd5cfa46ae2b72033e389236ddeb4","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-58dd2fe43e3d2dd60230b0cc","nextBlock":"b-1b722117caa5e008be58d3fb","prerequisites":[{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-1b722117caa5e008be58d3fb","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-1b722117caa5e008be58d3fb","content":"按问题继续","sha256":"e91f24e4e4ca40e2099a0d81d30db6568f4bf2074b499eeee6e3959617bd97ac","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-5e59d705a75c738ab1a89dd6","nextBlock":"b-9fc9426c423d234ee83ee5c8","prerequisites":[{"id":"b-5e59d705a75c738ab1a89dd6","url":"#b-5e59d705a75c738ab1a89dd6","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-9fc9426c423d234ee83ee5c8","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-9fc9426c423d234ee83ee5c8","content":"坐标、锚框偏移、训练/评价匹配、NMS/AP和真实检测训练连接","sha256":"88216e858779670a15bd42f20a06f136275ceee6a8685216f5821af4b1faf48f","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-1b722117caa5e008be58d3fb","nextBlock":"b-d5541ce223b78d216f754a6a","prerequisites":[{"id":"b-5e59d705a75c738ab1a89dd6","url":"#b-5e59d705a75c738ab1a89dd6","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-d5541ce223b78d216f754a6a","type":"h2","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-d5541ce223b78d216f754a6a","content":"## NMS 与评价匹配解决两个不同的重复问题","sha256":"72d3c7834c3fafe18571e3ac3fc55ec05967604ca8e1c604fc52313b1c6b033b","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-9fc9426c423d234ee83ee5c8","nextBlock":"b-4cd607b6283ba152137aee5c","prerequisites":[{"id":"b-5e59d705a75c738ab1a89dd6","url":"#b-5e59d705a75c738ab1a89dd6","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-4cd607b6283ba152137aee5c","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-4cd607b6283ba152137aee5c","content":"Non-maximum suppression（NMS）是推理后处理：通常按类把预测框按分数排序，保留最高分框，删除与之重叠超过阈值的其他框，再继续。它减少同一对象的重复报告，但不知道真值；挤在一起的不同对象也可能被误删。NMS 阈值影响交付的结果，不能在测试集上追着最好 AP 挑阈值。","sha256":"b117cf446d2ad20730ef6f594c83df176d7660e6471a02cd3f1b71e3e15edf9e","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d5541ce223b78d216f754a6a","nextBlock":"b-d92e5777909882aadc531f81","prerequisites":[{"id":"b-d5541ce223b78d216f754a6a","url":"#b-d5541ce223b78d216f754a6a","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-d92e5777909882aadc531f81","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-d92e5777909882aadc531f81","content":"评价匹配则拿最终预测与真值比较。在固定类别和 IoU 阈值下，按置信度从高到低，每个预测至多匹配一个尚未被使用的真值；同一真值只能被命中一次。于是“框覆盖了对象”还不保证是真阳性，它可能是已有更高分预测的重复框。","sha256":"315a10d85c66d1a4cc67a41fd3d34c08e2b14631a03e3d0174dc9b1de08132f4","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-4cd607b6283ba152137aee5c","nextBlock":"b-afc4aee1cb13579c45453fb3","prerequisites":[{"id":"b-d5541ce223b78d216f754a6a","url":"#b-d5541ce223b78d216f754a6a","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-afc4aee1cb13579c45453fb3","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-afc4aee1cb13579c45453fb3","content":"设共有两个真值，四个预测排序后的身份依次为 TP、FP（重复）、FP（背景）、TP。累计 precision 为 $1,1/2,1/3,1/2$，recall 为 $1/2,1/2,1/2,1$。若采用本册的全点插值 precision envelope 与 recall 增量积分，AP 为 $0.5\\times1+0.5\\times0.5=0.75$。这是明确定义的简化 AP；COCO 的 101 个 recall 阈值采样、多个 IoU 阈值平均和面积分组是更具体的另一套协议，不能把一个数未经说明叫“COCO AP”。","sha256":"ac5d224648f4ebde30f06c9a67e260c09658b54cddb2f0ff694279ac674969bc","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[{"latex":"1,1/2,1/3,1/2","definitionContext":"NMS 与评价匹配解决两个不同的重复问题"},{"latex":"1/2,1/2,1/2,1","definitionContext":"NMS 与评价匹配解决两个不同的重复问题"},{"latex":"0.5\\times1+0.5\\times0.5=0.75","definitionContext":"NMS 与评价匹配解决两个不同的重复问题"}],"previousBlock":"b-d92e5777909882aadc531f81","nextBlock":"b-7e7918882942596f8b268e6b","prerequisites":[{"id":"b-d5541ce223b78d216f754a6a","url":"#b-d5541ce223b78d216f754a6a","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7e7918882942596f8b268e6b","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-7e7918882942596f8b268e6b","content":"[LAB 09](https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/90d213b4e1e04c7e889ee6fc042b2358?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=90d213b4e1e04c7e889ee6fc042b2358&utm_term=get_notebook) 可改重复框分数，观察排序如何改变 AP；也可改 IoU 阈值，使一个原先合格的框失去匹配。准确率把大量背景当作普通样本时可能很高，却无法体现检测的定位与重复问题。","sha256":"fdf90e10bf1db95f074762e097e80c61cd82e960b2ac2808385f306853065c3c","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-afc4aee1cb13579c45453fb3","nextBlock":"b-467c3a2fcc3349fc3731f705","prerequisites":[{"id":"b-d5541ce223b78d216f754a6a","url":"#b-d5541ce223b78d216f754a6a","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-467c3a2fcc3349fc3731f705","type":"h2","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-467c3a2fcc3349fc3731f705","content":"## 沿一次真实训练追踪从图像到框","sha256":"9f8de4572c912ab854fcccf5932a372018412e10c44dbb237cf7260d10cfc308","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7e7918882942596f8b268e6b","nextBlock":"b-f7d2cbee72c87aec06cd8b22","prerequisites":[{"id":"b-d5541ce223b78d216f754a6a","url":"#b-d5541ce223b78d216f754a6a","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-f7d2cbee72c87aec06cd8b22","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-f7d2cbee72c87aec06cd8b22","content":"[LAB 10](https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/a14237bfa9dd413d959ab4e458e11f42?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=a14237bfa9dd413d959ab4e458e11f42&utm_term=get_notebook) 的主实验应按以下对象检查：RGB 图像和类别/框标注进入数据集，卷积骨干产生空间特征，输出头产生类别与偏移，匹配提供监督，损失更新参数，解码及 NMS 产生最终框，最后仅用保留图像评价。真值坐标只用于目标与评价，不能作为图像模型的输入特征。","sha256":"79ae2f09f6174dcb8948392160784f39aefcad99854bef12aef60b9ed330c357","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-467c3a2fcc3349fc3731f705","nextBlock":"b-5b5fbb7ac469d758aeb27b88","prerequisites":[{"id":"b-467c3a2fcc3349fc3731f705","url":"#b-467c3a2fcc3349fc3731f705","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-5b5fbb7ac469d758aeb27b88","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-5b5fbb7ac469d758aeb27b88","content":"主实验已在 2026-09-22 本地 CPU 完整运行。输入是 48×48 RGB 合成图，含红矩形、蓝椭圆、可变位置与尺度、多个对象及空图；颜色与形状相关，因此不是自然场景通用检测基准。按原始场景种子划分 768 训练、96 验证、192 保留图。45,806 参数的卷积检测器在单层网格上使用 288 个锚框，训练 28 epoch；正例阈值 0.45、忽略区间 [0.25,0.45)，并保障每个真值有独立强制匹配。分类使用 hard-negative CE，位置项为正例 smooth-L1，解码后按类 NMS。","sha256":"939c4761628c5da65993b40a6d844bfb7378ef29dbeeb37b428d08c76b31f013","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f7d2cbee72c87aec06cd8b22","nextBlock":"b-dd5228970e460ce1a5b60017","prerequisites":[{"id":"b-467c3a2fcc3349fc3731f705","url":"#b-467c3a2fcc3349fc3731f705","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-dd5228970e460ce1a5b60017","type":"table","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-dd5228970e460ce1a5b60017","content":"| 同一保留协议 | 两类平均 AP@IoU0.50 | 两类平均 AP@IoU0.75 |\n| --- | --- | --- |\n| 随机初始化 | 0.00746 | 0.00003 |\n| 训练后，原尺度 | 0.97303 | 0.70367 |\n| 训练后，仅对象尺寸减半 | 0.61490 | 0.17987 |","sha256":"f24808c9a952fb28fdcdc02be7ca78258ce96c4f5df42045c6c1f671959e3cb0","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"table":{"headers":["同一保留协议","两类平均 AP@IoU0.50","两类平均 AP@IoU0.75"],"rows":[["同一保留协议","两类平均 AP@IoU0.50","两类平均 AP@IoU0.75"],["随机初始化","0.00746","0.00003"],["训练后，原尺度","0.97303","0.70367"],["训练后，仅对象尺寸减半","0.61490","0.17987"]]},"previousBlock":"b-5b5fbb7ac469d758aeb27b88","nextBlock":"b-67205fa5b0862ac5f951d0d3","prerequisites":[{"id":"b-467c3a2fcc3349fc3731f705","url":"#b-467c3a2fcc3349fc3731f705","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-67205fa5b0862ac5f951d0d3","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-67205fa5b0862ac5f951d0d3","content":"AP 按各类全局分数排序、一个真值最多匹配一次、precision envelope 全点积分；不是 COCO 多 IoU 平均 AP。保留集有 33 张空图，展示阈值 0.4 下没有空图误报；这不代表所有阈值或所有背景都不会误报。AP50 与 AP75 的差距说明定位仍不够精细，尺度减半的下降说明训练分布外的小对象尤其困难。","sha256":"8a25829866cb20b99f921364d86c33657dae9e1547549d9560d15ceb1a6f9537","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-dd5228970e460ce1a5b60017","nextBlock":"b-9864a3ae81d5531e1d91a1d1","prerequisites":[{"id":"b-467c3a2fcc3349fc3731f705","url":"#b-467c3a2fcc3349fc3731f705","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-9864a3ae81d5531e1d91a1d1","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-9864a3ae81d5531e1d91a1d1","content":"![真实模型的保留预测与失败](/assets/display/88df211a98edfb7b00515441.webp)","sha256":"09e76a6115a60db2f769dcaf9677a8e9e048c037ec149a8c650a05d73a12af9f","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-67205fa5b0862ac5f951d0d3","nextBlock":"b-0857a6d2d9beea2f1cdc55ad","prerequisites":[{"id":"b-467c3a2fcc3349fc3731f705","url":"#b-467c3a2fcc3349fc3731f705","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-0857a6d2d9beea2f1cdc55ad","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-0857a6d2d9beea2f1cdc55ad","content":"![仅对象尺寸减半时的预测](/assets/display/735d18643da4042bf639ef2c.webp)","sha256":"9ab6840d67cc4c3ac9ff62f6ec9a8c36064f02605214b40349dbf9191fcbd589","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-9864a3ae81d5531e1d91a1d1","nextBlock":"b-19733dca83a79fd874714af1","prerequisites":[{"id":"b-467c3a2fcc3349fc3731f705","url":"#b-467c3a2fcc3349fc3731f705","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-19733dca83a79fd874714af1","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-19733dca83a79fd874714af1","content":"保持权重与输入不变，仅把 NMS IoU 阈值从 0.35 放宽到 0.65，AP50 变为 0.95866；展示阈值下 recall 从 0.98311 到 1.0，precision 则从 0.94175 降到 0.41926。更多预测覆盖了剩余真值，也留下大量重复框。这组结果把 NMS 的取舍落实到了同一模型，而非抽象宣称后处理总会改善成绩。","sha256":"d50b94eb722dee5ced832487e587531ad8fc0c7e1ea6a13ff5fe4d0906e93d7b","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-0857a6d2d9beea2f1cdc55ad","nextBlock":"b-42389b57dbaa0d6b5d24f7b4","prerequisites":[{"id":"b-467c3a2fcc3349fc3731f705","url":"#b-467c3a2fcc3349fc3731f705","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-42389b57dbaa0d6b5d24f7b4","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-42389b57dbaa0d6b5d24f7b4","content":"四维高斯样本上的逻辑回归仍能说明 sigmoid 分类损失和参数更新，但它没有图像和位置输出，应保留为分类热身。完整检测器给出的错误则可以进一步分解：漏检、错类、框偏、重复、空图误报、尺度外推。这些区分决定下一步该改表示、匹配、损失、解码还是数据。","sha256":"8b7d63701e0e921bee18b7f1c1e38d6db5d9a5db3cb2e40577eaa14a66c80021","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-19733dca83a79fd874714af1","nextBlock":"b-725960febe596dd5ec241fd8","prerequisites":[{"id":"b-467c3a2fcc3349fc3731f705","url":"#b-467c3a2fcc3349fc3731f705","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-725960febe596dd5ec241fd8","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-725960febe596dd5ec241fd8","content":"即使检测框全部正确，框里的杯把空洞和两杯接触边界仍未知。下一册 [CLS 08](https://deepnote.com/workspace/ise3309jingao-5ba66b0a-4283-4f5a-ae06-3f68d3201e99/project/38755e4e-23f6-4f5f-8c68-8a2e4114b1c1/notebook/2d81e2a1936043349eacb71c53c0defa?utm_source=openai&utm_medium=mcp&utm_campaign=openaimcp&utm_content=2d81e2a1936043349eacb71c53c0defa&utm_term=get_notebook) 从这个具体缺口走向逐像素和实例输出。","sha256":"47399423a3cd722f6cc81989106117ca9567f5f15f0b56d84fbad77bbc4db80c","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-opening","title":"NMS 与评价匹配解决两个不同的重复问题","number":1},"sectionTitle":"NMS 与评价匹配解决两个不同的重复问题","currentQuestion":"NMS 与评价匹配解决两个不同的重复问题","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-42389b57dbaa0d6b5d24f7b4","nextBlock":null,"prerequisites":[{"id":"b-467c3a2fcc3349fc3731f705","url":"#b-467c3a2fcc3349fc3731f705","role":"section-title"},{"id":"b-58dd2fe43e3d2dd60230b0cc","url":"#b-58dd2fe43e3d2dd60230b0cc","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-214b04ddaa0b8da05dfef01e","type":"h4","scope":"reading-section-1","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-214b04ddaa0b8da05dfef01e","content":"#### 回看：共享卷积、RPN 与一阶段检测的计算位置","sha256":"9d01ec975dd28a25f7030eff6e9ed5dac5ce031ab871645fb164d3f3b8ed1dfa","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-section-1","title":"回看：共享卷积、RPN 与一阶段检测的计算位置","number":2},"sectionTitle":"回看：共享卷积、RPN 与一阶段检测的计算位置","currentQuestion":"回看：共享卷积、RPN 与一阶段检测的计算位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-f8fc7263b378ee3f981acb51","prerequisites":[{"id":"b-505e617fd7c6420f7c20ecfd","url":"#b-505e617fd7c6420f7c20ecfd","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-f8fc7263b378ee3f981acb51","type":"h2","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-f8fc7263b378ee3f981acb51","content":"## 共享特征减少重复计算，但对象身份仍要保留","sha256":"11cc60ecac5230cad0208d185147fdee5e90412fc3a83335039e6be3103df597","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-section-1","title":"回看：共享卷积、RPN 与一阶段检测的计算位置","number":2},"sectionTitle":"回看：共享卷积、RPN 与一阶段检测的计算位置","currentQuestion":"回看：共享卷积、RPN 与一阶段检测的计算位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-214b04ddaa0b8da05dfef01e","nextBlock":"b-505e617fd7c6420f7c20ecfd","prerequisites":[{"id":"b-214b04ddaa0b8da05dfef01e","url":"#b-214b04ddaa0b8da05dfef01e","role":"section-title"},{"id":"b-505e617fd7c6420f7c20ecfd","url":"#b-505e617fd7c6420f7c20ecfd","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-505e617fd7c6420f7c20ecfd","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-505e617fd7c6420f7c20ecfd","content":"早期区域 CNN 为很多候选分别提特征，重复处理重叠像素。Fast R-CNN 把卷积移到整图共享阶段，再在区域内读取特征；Faster R-CNN 用 RPN 从共享特征提出候选；SSD 在多个尺度的特征图上直接预测类别和位置。这些变化围绕候选产生、特征复用和预测阶段数组织，并不意味着所有方法在任何设备或目标尺度上只有一个速度排序。[Fast R-CNN](https://arxiv.org/abs/1504.08083) · [Faster R-CNN](https://arxiv.org/abs/1506.01497) · [SSD](https://arxiv.org/abs/1512.02325)。","sha256":"bee32e87d504e0167d898e9c0f91a56e4c7b091c9b8feb49d487de4dab842691","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-section-1","title":"回看：共享卷积、RPN 与一阶段检测的计算位置","number":2},"sectionTitle":"回看：共享卷积、RPN 与一阶段检测的计算位置","currentQuestion":"回看：共享卷积、RPN 与一阶段检测的计算位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f8fc7263b378ee3f981acb51","nextBlock":"b-f42c34c02f11c1ec409055b0","prerequisites":[{"id":"b-f8fc7263b378ee3f981acb51","url":"#b-f8fc7263b378ee3f981acb51","role":"section-title"}],"relatedBlocks":[]},{"id":"b-f42c34c02f11c1ec409055b0","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-f42c34c02f11c1ec409055b0","content":"共享整图特征不应抹掉区域身份。同一张特征图上取左杯和右杯两个区域，仍要得到各自预测；把所有空间位置先平均成一个向量，再期待检测头重建任意细节，可能遇到表示碰撞。是否保留坐标、感受野与采样密度，应与表示的分类能力一起检查。","sha256":"4746f9c7084f10c1639a38d91b9ad149f85aba184e86b9c42f324db145121c64","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-section-1","title":"回看：共享卷积、RPN 与一阶段检测的计算位置","number":2},"sectionTitle":"回看：共享卷积、RPN 与一阶段检测的计算位置","currentQuestion":"回看：共享卷积、RPN 与一阶段检测的计算位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-505e617fd7c6420f7c20ecfd","nextBlock":"b-028f808e19496034f92c0528","prerequisites":[{"id":"b-f8fc7263b378ee3f981acb51","url":"#b-f8fc7263b378ee3f981acb51","role":"section-title"},{"id":"b-505e617fd7c6420f7c20ecfd","url":"#b-505e617fd7c6420f7c20ecfd","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-028f808e19496034f92c0528","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f.html#b-028f808e19496034f92c0528","content":"[一阶段检测如何对应 LAB 10](/course/blocks/dense-detector-ssd-extension.html)","sha256":"609abc3860560a383fa699b0b71970390c0f3d0f80d59952beb64f4d04b0862a","chapter":{"id":"course","title":"NMS 与评价匹配解决两个不同的重复问题"},"unit":{"id":"notebook-blocks/8f46ef44ed4e4529beedf33817f14d3f","title":"NMS 与评价匹配解决两个不同的重复问题"},"page":{"id":"reading-section-1","title":"回看：共享卷积、RPN 与一阶段检测的计算位置","number":2},"sectionTitle":"回看：共享卷积、RPN 与一阶段检测的计算位置","currentQuestion":"回看：共享卷积、RPN 与一阶段检测的计算位置","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f42c34c02f11c1ec409055b0","nextBlock":null,"prerequisites":[{"id":"b-f8fc7263b378ee3f981acb51","url":"#b-f8fc7263b378ee3f981acb51","role":"section-title"},{"id":"b-505e617fd7c6420f7c20ecfd","url":"#b-505e617fd7c6420f7c20ecfd","role":"problem-context"}],"relatedBlocks":[]}]}