{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/blocks/roi-align-paper-evidence.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"原论文的对照：对齐改善了哪些指标","url":"https://codingai-lec03.pages.dev/course/blocks/roi-align-paper-evidence.html","markdown":"https://codingai-lec03.pages.dev/course/blocks/roi-align-paper-evidence.md","pages":[{"id":"reading-opening","title":"原论文的对照：对齐改善了哪些指标","unit":"区域采样怎样保住几何对应？","unitId":"roi-alignment-and-evidence","chapter":"II","number":1}],"aliases":[{"oldId":"b-a333dda09670760662f71798","targetId":"b-a333dda09670760662f71798","page":"reading-opening","operation":"move","beforeSHA256":"b97c7fde41ad013173ed390a156fe63c7df50ad0df13e7e377bd5b103516e9ce","afterSHA256":"b97c7fde41ad013173ed390a156fe63c7df50ad0df13e7e377bd5b103516e9ce","reason":"Same source body in authored semantic page"},{"oldId":"b-9a003fd0387172fe119e5f57","targetId":"b-9a003fd0387172fe119e5f57","page":"reading-opening","operation":"move","beforeSHA256":"185d6570689468f3c523e1d63cc5d2dd507c9b0ff60f2af36bd85b2834d7978e","afterSHA256":"185d6570689468f3c523e1d63cc5d2dd507c9b0ff60f2af36bd85b2834d7978e","reason":"Same source body in authored semantic page"},{"oldId":"b-449a648020154909d0f3a016","targetId":"b-449a648020154909d0f3a016","page":"reading-opening","operation":"move","beforeSHA256":"3bb8d279d8595e2246936d624e3b4edea6ceca91fece02cd289fada770a219f4","afterSHA256":"3bb8d279d8595e2246936d624e3b4edea6ceca91fece02cd289fada770a219f4","reason":"Same source body in authored semantic page"},{"oldId":"b-74bc10ea753c736215f2fc26","targetId":"b-74bc10ea753c736215f2fc26","page":"reading-opening","operation":"rewrite","beforeSHA256":"0bfedb1ed97588d37565883224ab357ea9af5995f6404d1f3d1042d261fbdfd8","afterSHA256":"f59df6ce113db0faf2717e5b8440d85ef1f069b2e8b327b430459828cfd41119","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-dcc353de3ea38a3e7a05180d","targetId":"b-dcc353de3ea38a3e7a05180d","page":"reading-opening","operation":"move","beforeSHA256":"4cdba8cde1b8262a042dd6aa8cda6abe7fd5a4e8cb6f224f5bb95e0f3611cd8d","afterSHA256":"4cdba8cde1b8262a042dd6aa8cda6abe7fd5a4e8cb6f224f5bb95e0f3611cd8d","reason":"Same source body in authored semantic page"},{"oldId":"b-b1da328de11b7c990b7bd616","targetId":"b-b1da328de11b7c990b7bd616","page":"reading-opening","operation":"move","beforeSHA256":"a75c59c916a80252869b500277c7120d64fcb5056b16b7893076a6f49e7c3632","afterSHA256":"a75c59c916a80252869b500277c7120d64fcb5056b16b7893076a6f49e7c3632","reason":"Same source body in authored semantic page"},{"oldId":"b-7196a75d6bfb3e22a3b9ef1f","targetId":"b-7196a75d6bfb3e22a3b9ef1f","page":"reading-opening","operation":"move","beforeSHA256":"50177d63e411d3e58553d7d266c6c7aa4db901ec7404aeb2df04d4279b19a9b5","afterSHA256":"50177d63e411d3e58553d7d266c6c7aa4db901ec7404aeb2df04d4279b19a9b5","reason":"Same source body in authored semantic page"},{"oldId":"b-6e52a42e08007cfc6bdeb254","targetId":"b-6e52a42e08007cfc6bdeb254","page":"reading-opening","operation":"move","beforeSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","afterSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","reason":"Same source body in authored semantic page"}],"blocks":[{"id":"b-a333dda09670760662f71798","type":"h1","scope":"roi-align-paper-evidence","url":"https://codingai-lec03.pages.dev/course/blocks/roi-align-paper-evidence.html#b-a333dda09670760662f71798","content":"# 原论文的对照：对齐改善了哪些指标","sha256":"b97c7fde41ad013173ed390a156fe63c7df50ad0df13e7e377bd5b103516e9ce","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"roi-alignment-and-evidence","title":"区域采样怎样保住几何对应？"},"page":{"id":"reading-opening","title":"原论文的对照：对齐改善了哪些指标","number":1},"sectionTitle":"原论文的对照：对齐改善了哪些指标","currentQuestion":"原论文的对照：对齐改善了哪些指标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-9a003fd0387172fe119e5f57","prerequisites":[{"id":"b-9a003fd0387172fe119e5f57","url":"#b-9a003fd0387172fe119e5f57","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-9a003fd0387172fe119e5f57","type":"p","scope":"roi-align-paper-evidence","url":"https://codingai-lec03.pages.dev/course/blocks/roi-align-paper-evidence.html#b-9a003fd0387172fe119e5f57","content":"对 0、2、4、6 四邻值，按横向 1/4、纵向 1/2 算出 2.5。两步插值与四项加权是同一运算。读完算例后，接着看[原论文的对照：对齐改善了哪些指标](/course/blocks/roi-align-paper-evidence.html)，区分“使用双线性插值”和“保持坐标对齐”的作用。","sha256":"185d6570689468f3c523e1d63cc5d2dd507c9b0ff60f2af36bd85b2834d7978e","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"roi-alignment-and-evidence","title":"区域采样怎样保住几何对应？"},"page":{"id":"reading-opening","title":"原论文的对照：对齐改善了哪些指标","number":1},"sectionTitle":"原论文的对照：对齐改善了哪些指标","currentQuestion":"原论文的对照：对齐改善了哪些指标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-a333dda09670760662f71798","nextBlock":"b-449a648020154909d0f3a016","prerequisites":[{"id":"b-a333dda09670760662f71798","url":"#b-a333dda09670760662f71798","role":"section-title"}],"relatedBlocks":[]},{"id":"b-449a648020154909d0f3a016","type":"p","scope":"roi-align-paper-evidence","url":"https://codingai-lec03.pages.dev/course/blocks/roi-align-paper-evidence.html#b-449a648020154909d0f3a016","content":"Mask R-CNN v3 的 Table 2 使用 COCO `trainval35k` 训练、`minival` 测试；表中未另说明时报告 mask AP。以下固定 ResNet-50-C4、stride 16，并选同为 max 聚合的行：","sha256":"3bb8d279d8595e2246936d624e3b4edea6ceca91fece02cd289fada770a219f4","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"roi-alignment-and-evidence","title":"区域采样怎样保住几何对应？"},"page":{"id":"reading-opening","title":"原论文的对照：对齐改善了哪些指标","number":1},"sectionTitle":"原论文的对照：对齐改善了哪些指标","currentQuestion":"原论文的对照：对齐改善了哪些指标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-9a003fd0387172fe119e5f57","nextBlock":"b-74bc10ea753c736215f2fc26","prerequisites":[{"id":"b-a333dda09670760662f71798","url":"#b-a333dda09670760662f71798","role":"section-title"},{"id":"b-9a003fd0387172fe119e5f57","url":"#b-9a003fd0387172fe119e5f57","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-74bc10ea753c736215f2fc26","type":"table","scope":"roi-align-paper-evidence","url":"https://codingai-lec03.pages.dev/course/blocks/roi-align-paper-evidence.html#b-74bc10ea753c736215f2fc26","content":"| 区域读取 | 是否保留对齐 | 是否双线性 | mask AP | mask AP75 |\n| --- | --- | --- | --- | --- |\n| RoIPool | 否 | 否 | 26.9 | 26.4 |\n| RoIWarp | 否 | 是 | 27.2 | 27.1 |\n| RoIAlign | 是 | 是 | 30.2 | 31.8 |","sha256":"f59df6ce113db0faf2717e5b8440d85ef1f069b2e8b327b430459828cfd41119","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"roi-alignment-and-evidence","title":"区域采样怎样保住几何对应？"},"page":{"id":"reading-opening","title":"原论文的对照：对齐改善了哪些指标","number":1},"sectionTitle":"原论文的对照：对齐改善了哪些指标","currentQuestion":"原论文的对照：对齐改善了哪些指标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"table":{"headers":["区域读取","是否保留对齐","是否双线性","mask AP","mask AP75"],"rows":[["区域读取","是否保留对齐","是否双线性","mask AP","mask AP75"],["RoIPool","否","否","26.9","26.4"],["RoIWarp","否","是","27.2","27.1"],["RoIAlign","是","是","30.2","31.8"]]},"previousBlock":"b-449a648020154909d0f3a016","nextBlock":"b-dcc353de3ea38a3e7a05180d","prerequisites":[{"id":"b-a333dda09670760662f71798","url":"#b-a333dda09670760662f71798","role":"section-title"},{"id":"b-9a003fd0387172fe119e5f57","url":"#b-9a003fd0387172fe119e5f57","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-dcc353de3ea38a3e7a05180d","type":"p","scope":"roi-align-paper-evidence","url":"https://codingai-lec03.pages.dev/course/blocks/roi-align-paper-evidence.html#b-dcc353de3ea38a3e7a05180d","content":"RoIAlign 相比 RoIPool 的 AP 增加 3.3 点，AP75 增加 5.4 点。RoIWarp 也插值却没有同样改善，支持“保持坐标对应”这一机制判断。平均聚合的 RoIAlign 行是 AP30.3/AP75 31.5，引用时不能把不同聚合行混在一个对照里。Table 2(d) 换成 C5、stride32，mask AP 从 23.6 到 30.9；这是另一组设置。","sha256":"4cdba8cde1b8262a042dd6aa8cda6abe7fd5a4e8cb6f224f5bb95e0f3611cd8d","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"roi-alignment-and-evidence","title":"区域采样怎样保住几何对应？"},"page":{"id":"reading-opening","title":"原论文的对照：对齐改善了哪些指标","number":1},"sectionTitle":"原论文的对照：对齐改善了哪些指标","currentQuestion":"原论文的对照：对齐改善了哪些指标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-74bc10ea753c736215f2fc26","nextBlock":"b-b1da328de11b7c990b7bd616","prerequisites":[{"id":"b-a333dda09670760662f71798","url":"#b-a333dda09670760662f71798","role":"section-title"},{"id":"b-9a003fd0387172fe119e5f57","url":"#b-9a003fd0387172fe119e5f57","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-b1da328de11b7c990b7bd616","type":"p","scope":"roi-align-paper-evidence","url":"https://codingai-lec03.pages.dev/course/blocks/roi-align-paper-evidence.html#b-b1da328de11b7c990b7bd616","content":"这些是训练完整模型的论文结果，不是本课程重跑。它们说明读取位置值得修复，尤其严格定位受益；P44 则把输入 F 固定成完全相同，讨论任何确定性读出都无法创造差别。方法有效与构造例中的信息限制，可以同时成立。","sha256":"a75c59c916a80252869b500277c7120d64fcb5056b16b7893076a6f49e7c3632","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"roi-alignment-and-evidence","title":"区域采样怎样保住几何对应？"},"page":{"id":"reading-opening","title":"原论文的对照：对齐改善了哪些指标","number":1},"sectionTitle":"原论文的对照：对齐改善了哪些指标","currentQuestion":"原论文的对照：对齐改善了哪些指标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-dcc353de3ea38a3e7a05180d","nextBlock":"b-7196a75d6bfb3e22a3b9ef1f","prerequisites":[{"id":"b-a333dda09670760662f71798","url":"#b-a333dda09670760662f71798","role":"section-title"},{"id":"b-9a003fd0387172fe119e5f57","url":"#b-9a003fd0387172fe119e5f57","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7196a75d6bfb3e22a3b9ef1f","type":"p","scope":"roi-align-paper-evidence","url":"https://codingai-lec03.pages.dev/course/blocks/roi-align-paper-evidence.html#b-7196a75d6bfb3e22a3b9ef1f","content":"接着读[为什么相同 F 仍无法恢复：表示碰撞的推导](/course/blocks/representation-collision-proof.html)，再到 [LAB03 · 空间信息：表示、采样与掩码](/course/notebooks/49a2c030eafa4e1db6fc4f1a08c3810a.html)，用一维读数和差分通道改变条件。","sha256":"50177d63e411d3e58553d7d266c6c7aa4db901ec7404aeb2df04d4279b19a9b5","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"roi-alignment-and-evidence","title":"区域采样怎样保住几何对应？"},"page":{"id":"reading-opening","title":"原论文的对照：对齐改善了哪些指标","number":1},"sectionTitle":"原论文的对照：对齐改善了哪些指标","currentQuestion":"原论文的对照：对齐改善了哪些指标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b1da328de11b7c990b7bd616","nextBlock":"b-6e52a42e08007cfc6bdeb254","prerequisites":[{"id":"b-a333dda09670760662f71798","url":"#b-a333dda09670760662f71798","role":"section-title"},{"id":"b-9a003fd0387172fe119e5f57","url":"#b-9a003fd0387172fe119e5f57","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-6e52a42e08007cfc6bdeb254","type":"h2","scope":"roi-align-paper-evidence","url":"https://codingai-lec03.pages.dev/course/blocks/roi-align-paper-evidence.html#b-6e52a42e08007cfc6bdeb254","content":"## 继续阅读与实验","sha256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"roi-alignment-and-evidence","title":"区域采样怎样保住几何对应？"},"page":{"id":"reading-opening","title":"原论文的对照：对齐改善了哪些指标","number":1},"sectionTitle":"原论文的对照：对齐改善了哪些指标","currentQuestion":"原论文的对照：对齐改善了哪些指标","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7196a75d6bfb3e22a3b9ef1f","nextBlock":null,"prerequisites":[{"id":"b-a333dda09670760662f71798","url":"#b-a333dda09670760662f71798","role":"section-title"},{"id":"b-9a003fd0387172fe119e5f57","url":"#b-9a003fd0387172fe119e5f57","role":"problem-context"}],"relatedBlocks":[]}]}