{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/blocks/small-object-confusion.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"像素准确率很高，为什么小对象仍可能很差","url":"https://codingai-lec03.pages.dev/course/blocks/small-object-confusion.html","markdown":"https://codingai-lec03.pages.dev/course/blocks/small-object-confusion.md","pages":[{"id":"reading-opening","title":"像素准确率很高，为什么小对象仍可能很差","unit":"哪些信息真的改善了像素预测？","unitId":"pixel-detail-and-baselines","chapter":"II","number":1}],"aliases":[{"oldId":"b-11f931dd70a46f739199c9cf","targetId":"b-11f931dd70a46f739199c9cf","page":"reading-opening","operation":"move","beforeSHA256":"2ae6ad4b71b95ea6128afe0f16444fc82574b76e78555d1ec57c29ca22b1b161","afterSHA256":"2ae6ad4b71b95ea6128afe0f16444fc82574b76e78555d1ec57c29ca22b1b161","reason":"Same source body in authored semantic page"},{"oldId":"b-a93beea7dcaddfea37fd9654","targetId":"b-a93beea7dcaddfea37fd9654","page":"reading-opening","operation":"move","beforeSHA256":"6dd29e183c85374a7e5328c1f14612c2727e29f545b1886fc3c97b1ecc5ff3dd","afterSHA256":"6dd29e183c85374a7e5328c1f14612c2727e29f545b1886fc3c97b1ecc5ff3dd","reason":"Same source body in authored semantic page"},{"oldId":"b-3d39cd370cd1f96aa71ec4aa","targetId":"b-3d39cd370cd1f96aa71ec4aa","page":"reading-opening","operation":"move","beforeSHA256":"23ffd6b88ea76fcd50eb098e2dc963b89205c90a9dce967db318c27a9cbd32bc","afterSHA256":"23ffd6b88ea76fcd50eb098e2dc963b89205c90a9dce967db318c27a9cbd32bc","reason":"Same source body in authored semantic page"},{"oldId":"b-99b308c7391d851d78299d25","targetId":"b-99b308c7391d851d78299d25","page":"reading-opening","operation":"move","beforeSHA256":"ac6c96517ea38a8b31ee87244b5d36bbe93b4c986e4bf524838999c339aeb6a8","afterSHA256":"ac6c96517ea38a8b31ee87244b5d36bbe93b4c986e4bf524838999c339aeb6a8","reason":"Same source body in authored semantic page"},{"oldId":"b-d52ec21ff5791ab44660f9bb","targetId":"b-d52ec21ff5791ab44660f9bb","page":"reading-opening","operation":"move","beforeSHA256":"4f61043660ef3b6160569b139ba49eed3c45d92cbbd88ec18c30e5fad0bf5344","afterSHA256":"4f61043660ef3b6160569b139ba49eed3c45d92cbbd88ec18c30e5fad0bf5344","reason":"Same source body in authored semantic page"},{"oldId":"b-09e439f1fcc84dfdf2192c2d","targetId":"b-09e439f1fcc84dfdf2192c2d","page":"reading-opening","operation":"move","beforeSHA256":"e32593333f5cb0f17ff6b0e4c82151b539433bac13f141a144fd34a8ffa2cefa","afterSHA256":"e32593333f5cb0f17ff6b0e4c82151b539433bac13f141a144fd34a8ffa2cefa","reason":"Same source body in authored semantic page"},{"oldId":"b-c32fad7977415a16c370618f","targetId":"b-c32fad7977415a16c370618f","page":"reading-opening","operation":"move","beforeSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","afterSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","reason":"Same source body in authored semantic page"}],"blocks":[{"id":"b-11f931dd70a46f739199c9cf","type":"h1","scope":"small-object-confusion","url":"https://codingai-lec03.pages.dev/course/blocks/small-object-confusion.html#b-11f931dd70a46f739199c9cf","content":"# 像素准确率很高，为什么小对象仍可能很差","sha256":"2ae6ad4b71b95ea6128afe0f16444fc82574b76e78555d1ec57c29ca22b1b161","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"reading-opening","title":"像素准确率很高，为什么小对象仍可能很差","number":1},"sectionTitle":"像素准确率很高，为什么小对象仍可能很差","currentQuestion":"像素准确率很高，为什么小对象仍可能很差","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-a93beea7dcaddfea37fd9654","prerequisites":[{"id":"b-a93beea7dcaddfea37fd9654","url":"#b-a93beea7dcaddfea37fd9654","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-a93beea7dcaddfea37fd9654","type":"p","scope":"small-object-confusion","url":"https://codingai-lec03.pages.dev/course/blocks/small-object-confusion.html#b-a93beea7dcaddfea37fd9654","content":"较细支路给解码器另一份可能保留位置线索的输入，但不保证边界正确；应在同一任务比较错误图与逐类 IoU。现有实验中，RGB 简单基线在原条件下胜出，FCN 在模糊后更稳健。对照两种条件，判断结构何时实际有用。","sha256":"6dd29e183c85374a7e5328c1f14612c2727e29f545b1886fc3c97b1ecc5ff3dd","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"reading-opening","title":"像素准确率很高，为什么小对象仍可能很差","number":1},"sectionTitle":"像素准确率很高，为什么小对象仍可能很差","currentQuestion":"像素准确率很高，为什么小对象仍可能很差","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-11f931dd70a46f739199c9cf","nextBlock":"b-3d39cd370cd1f96aa71ec4aa","prerequisites":[{"id":"b-11f931dd70a46f739199c9cf","url":"#b-11f931dd70a46f739199c9cf","role":"section-title"}],"relatedBlocks":[]},{"id":"b-3d39cd370cd1f96aa71ec4aa","type":"p","scope":"small-object-confusion","url":"https://codingai-lec03.pages.dev/course/blocks/small-object-confusion.html#b-3d39cd370cd1f96aa71ec4aa","content":"一张大部分为背景的图，只要把背景判对，整体像素准确率就会很高。假设 10000 个像素里只有 100 个属于目标，全部预测背景就有 99% 准确率，但目标 IoU 为零。这个算例说明面积不平衡如何隐藏感兴趣类别的失败。","sha256":"23ffd6b88ea76fcd50eb098e2dc963b89205c90a9dce967db318c27a9cbd32bc","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"reading-opening","title":"像素准确率很高，为什么小对象仍可能很差","number":1},"sectionTitle":"像素准确率很高，为什么小对象仍可能很差","currentQuestion":"像素准确率很高，为什么小对象仍可能很差","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-a93beea7dcaddfea37fd9654","nextBlock":"b-99b308c7391d851d78299d25","prerequisites":[{"id":"b-11f931dd70a46f739199c9cf","url":"#b-11f931dd70a46f739199c9cf","role":"section-title"},{"id":"b-a93beea7dcaddfea37fd9654","url":"#b-a93beea7dcaddfea37fd9654","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-99b308c7391d851d78299d25","type":"p","scope":"small-object-confusion","url":"https://codingai-lec03.pages.dev/course/blocks/small-object-confusion.html#b-99b308c7391d851d78299d25","content":"LAB11 的同权重尺度 ×.45 对照更具体：像素准确率 .99816，而蓝椭圆 IoU 为 .85763、红矩形 .97613、背景 .99812。按三类等权平均得到 .94396。蓝对象与背景窄缝更细，发生少量像素错误就显著影响目标的比例。","sha256":"ac6c96517ea38a8b31ee87244b5d36bbe93b4c986e4bf524838999c339aeb6a8","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"reading-opening","title":"像素准确率很高，为什么小对象仍可能很差","number":1},"sectionTitle":"像素准确率很高，为什么小对象仍可能很差","currentQuestion":"像素准确率很高，为什么小对象仍可能很差","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-3d39cd370cd1f96aa71ec4aa","nextBlock":"b-d52ec21ff5791ab44660f9bb","prerequisites":[{"id":"b-11f931dd70a46f739199c9cf","url":"#b-11f931dd70a46f739199c9cf","role":"section-title"},{"id":"b-a93beea7dcaddfea37fd9654","url":"#b-a93beea7dcaddfea37fd9654","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-d52ec21ff5791ab44660f9bb","type":"p","scope":"small-object-confusion","url":"https://codingai-lec03.pages.dev/course/blocks/small-object-confusion.html#b-d52ec21ff5791ab44660f9bb","content":"查看失败图时要把输入、真值和错误像素并排。若杯把空洞消失，可能是原观测不清、特征混合、读出不对齐或任务头没学好；单看总数不能定位原因。可以先冻结其余条件，改变输入尺度或采样方式，再追踪错误是否沿边界集中。","sha256":"4f61043660ef3b6160569b139ba49eed3c45d92cbbd88ec18c30e5fad0bf5344","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"reading-opening","title":"像素准确率很高，为什么小对象仍可能很差","number":1},"sectionTitle":"像素准确率很高，为什么小对象仍可能很差","currentQuestion":"像素准确率很高，为什么小对象仍可能很差","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-99b308c7391d851d78299d25","nextBlock":"b-09e439f1fcc84dfdf2192c2d","prerequisites":[{"id":"b-11f931dd70a46f739199c9cf","url":"#b-11f931dd70a46f739199c9cf","role":"section-title"},{"id":"b-a93beea7dcaddfea37fd9654","url":"#b-a93beea7dcaddfea37fd9654","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-09e439f1fcc84dfdf2192c2d","type":"p","scope":"small-object-confusion","url":"https://codingai-lec03.pages.dev/course/blocks/small-object-confusion.html#b-09e439f1fcc84dfdf2192c2d","content":"本页连接细/粗特征的目的不是保证某个结构胜出，而是让我们可以问一个可测问题：增加细分支是否在同一输入和训练协议下，实际减少了所关心的窄缝错误？已有实验提供起点，完整架构对照仍需另设。","sha256":"e32593333f5cb0f17ff6b0e4c82151b539433bac13f141a144fd34a8ffa2cefa","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"reading-opening","title":"像素准确率很高，为什么小对象仍可能很差","number":1},"sectionTitle":"像素准确率很高，为什么小对象仍可能很差","currentQuestion":"像素准确率很高，为什么小对象仍可能很差","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d52ec21ff5791ab44660f9bb","nextBlock":"b-c32fad7977415a16c370618f","prerequisites":[{"id":"b-11f931dd70a46f739199c9cf","url":"#b-11f931dd70a46f739199c9cf","role":"section-title"},{"id":"b-a93beea7dcaddfea37fd9654","url":"#b-a93beea7dcaddfea37fd9654","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-c32fad7977415a16c370618f","type":"h2","scope":"small-object-confusion","url":"https://codingai-lec03.pages.dev/course/blocks/small-object-confusion.html#b-c32fad7977415a16c370618f","content":"## 继续阅读与实验","sha256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","chapter":{"id":"II","title":"区域、像素与实例"},"unit":{"id":"pixel-detail-and-baselines","title":"哪些信息真的改善了像素预测？"},"page":{"id":"reading-opening","title":"像素准确率很高，为什么小对象仍可能很差","number":1},"sectionTitle":"像素准确率很高，为什么小对象仍可能很差","currentQuestion":"像素准确率很高，为什么小对象仍可能很差","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-09e439f1fcc84dfdf2192c2d","nextBlock":null,"prerequisites":[{"id":"b-11f931dd70a46f739199c9cf","url":"#b-11f931dd70a46f739199c9cf","role":"section-title"},{"id":"b-a93beea7dcaddfea37fd9654","url":"#b-a93beea7dcaddfea37fd9654","role":"problem-context"}],"relatedBlocks":[]}]}