{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/blocks/grounding-pixel-interventions.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"已保存的指代实验：从像素提框后再应用关系","url":"https://codingai-lec03.pages.dev/course/blocks/grounding-pixel-interventions.html","markdown":"https://codingai-lec03.pages.dev/course/blocks/grounding-pixel-interventions.md","pages":[{"id":"reading-opening","title":"已保存的指代实验：从像素提框后再应用关系","unit":"一句话怎样落实为正确对象的区域？","unitId":"grounding-object-contract","chapter":"III","number":1}],"aliases":[{"oldId":"b-97d3bd79b5a7dadf94c8a8e4","targetId":"b-97d3bd79b5a7dadf94c8a8e4","page":"reading-opening","operation":"move","beforeSHA256":"b16fa63d28255b14ee6ad79feac2977745a35f5d22c9a83d9adddc2616d6e39c","afterSHA256":"b16fa63d28255b14ee6ad79feac2977745a35f5d22c9a83d9adddc2616d6e39c","reason":"Same source body in authored semantic page"},{"oldId":"b-bd15a843c17a310fed9ce2ad","targetId":"b-bd15a843c17a310fed9ce2ad","page":"reading-opening","operation":"move","beforeSHA256":"7d8bc1408f3f30ddf3278f32dbb54c475cf47f3b9b5ab00d4d7f1ef8f718627d","afterSHA256":"7d8bc1408f3f30ddf3278f32dbb54c475cf47f3b9b5ab00d4d7f1ef8f718627d","reason":"Same source body in authored semantic page"},{"oldId":"b-efb11cb39a4df88a2ec5a2b4","targetId":"b-efb11cb39a4df88a2ec5a2b4","page":"reading-opening","operation":"move","beforeSHA256":"ec817964628fe547da360e42935f3ab5e1d4ab0a413223243d9b9bdd08b46be5","afterSHA256":"ec817964628fe547da360e42935f3ab5e1d4ab0a413223243d9b9bdd08b46be5","reason":"Same source body in authored semantic page"},{"oldId":"b-32a3ea02da132a0d4c07182c","targetId":"b-32a3ea02da132a0d4c07182c","page":"reading-opening","operation":"rewrite","beforeSHA256":"73bea769aec3dcf143e6309aa7486e9552aae00c14ac4750b7956401686e39f3","afterSHA256":"0de466cf5672ac86c9dd4214b39ae872dd46ee32d4256b3291c661a7a495c361","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-37ddfcb7277dc75a119b9b58","targetId":"b-37ddfcb7277dc75a119b9b58","page":"reading-opening","operation":"move","beforeSHA256":"55fb5f3349686e772ea73d62b88bc243f3fdb3efd45df8c2454262f799c64329","afterSHA256":"55fb5f3349686e772ea73d62b88bc243f3fdb3efd45df8c2454262f799c64329","reason":"Same source body in authored semantic page"},{"oldId":"b-175046355e8512b5674f783a","targetId":"b-175046355e8512b5674f783a","page":"reading-opening","operation":"move","beforeSHA256":"0b5941126b2133f8855a3f934fc08044d64ac2370383c8794b2060a33380f994","afterSHA256":"70aab124a4de4ccb1629c768a54fc062777fd23a1d521943a55fc0d99ea993f7","reason":"Same source body in authored semantic page"},{"oldId":"b-01c0b1a977cb2249943ab387","targetId":"b-01c0b1a977cb2249943ab387","page":"reading-opening","operation":"move","beforeSHA256":"2c856e57e1c37c3d70093e2c5de7c9890f3a1bfe364eef8129588edb5dd7a622","afterSHA256":"2c856e57e1c37c3d70093e2c5de7c9890f3a1bfe364eef8129588edb5dd7a622","reason":"Same source body in authored semantic page"},{"oldId":"b-aec2bf693ad82a4e7e46cca0","targetId":"b-aec2bf693ad82a4e7e46cca0","page":"reading-opening","operation":"move","beforeSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","afterSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","reason":"Same source body in authored semantic page"}],"blocks":[{"id":"b-97d3bd79b5a7dadf94c8a8e4","type":"h1","scope":"grounding-pixel-interventions","url":"https://codingai-lec03.pages.dev/course/blocks/grounding-pixel-interventions.html#b-97d3bd79b5a7dadf94c8a8e4","content":"# 已保存的指代实验：从像素提框后再应用关系","sha256":"b16fa63d28255b14ee6ad79feac2977745a35f5d22c9a83d9adddc2616d6e39c","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"grounding-object-contract","title":"一句话怎样落实为正确对象的区域？"},"page":{"id":"reading-opening","title":"已保存的指代实验：从像素提框后再应用关系","number":1},"sectionTitle":"已保存的指代实验：从像素提框后再应用关系","currentQuestion":"已保存的指代实验：从像素提框后再应用关系","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-bd15a843c17a310fed9ce2ad","prerequisites":[{"id":"b-bd15a843c17a310fed9ce2ad","url":"#b-bd15a843c17a310fed9ce2ad","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-bd15a843c17a310fed9ce2ad","type":"p","scope":"grounding-pixel-interventions","url":"https://codingai-lec03.pages.dev/course/blocks/grounding-pixel-interventions.html#b-bd15a843c17a310fed9ce2ad","content":"语言条件落实为空间输出时，需要明确目标、参照物和关系。GLIP 提供对象级学习和迁移的研究路线；LAB15 提供可逐像素追踪的规则实验。两者的层级不同，但都帮助判断系统到底把框放在谁身上，再与后续掩码处理连接。","sha256":"7d8bc1408f3f30ddf3278f32dbb54c475cf47f3b9b5ab00d4d7f1ef8f718627d","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"grounding-object-contract","title":"一句话怎样落实为正确对象的区域？"},"page":{"id":"reading-opening","title":"已保存的指代实验：从像素提框后再应用关系","number":1},"sectionTitle":"已保存的指代实验：从像素提框后再应用关系","currentQuestion":"已保存的指代实验：从像素提框后再应用关系","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-97d3bd79b5a7dadf94c8a8e4","nextBlock":"b-efb11cb39a4df88a2ec5a2b4","prerequisites":[{"id":"b-97d3bd79b5a7dadf94c8a8e4","url":"#b-97d3bd79b5a7dadf94c8a8e4","role":"section-title"}],"relatedBlocks":[]},{"id":"b-efb11cb39a4df88a2ec5a2b4","type":"p","scope":"grounding-pixel-interventions","url":"https://codingai-lec03.pages.dev/course/blocks/grounding-pixel-interventions.html#b-efb11cb39a4df88a2ec5a2b4","content":"LAB15 的当前实现先从 RGB 图像按颜色阈值提取连通域，得到候选框；再解析目标、参照物与关系。它是真正读取像素的规则系统，类别由有限颜色约定提供，没有训练神经 grounding 模型。这个身份既不应被贬为“没读图”，也不应被升级为“复现 GLIP”。","sha256":"ec817964628fe547da360e42935f3ab5e1d4ab0a413223243d9b9bdd08b46be5","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"grounding-object-contract","title":"一句话怎样落实为正确对象的区域？"},"page":{"id":"reading-opening","title":"已保存的指代实验：从像素提框后再应用关系","number":1},"sectionTitle":"已保存的指代实验：从像素提框后再应用关系","currentQuestion":"已保存的指代实验：从像素提框后再应用关系","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-bd15a843c17a310fed9ce2ad","nextBlock":"b-32a3ea02da132a0d4c07182c","prerequisites":[{"id":"b-97d3bd79b5a7dadf94c8a8e4","url":"#b-97d3bd79b5a7dadf94c8a8e4","role":"section-title"},{"id":"b-bd15a843c17a310fed9ce2ad","url":"#b-bd15a843c17a310fed9ce2ad","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-32a3ea02da132a0d4c07182c","type":"table","scope":"grounding-pixel-interventions","url":"https://codingai-lec03.pages.dev/course/blocks/grounding-pixel-interventions.html#b-32a3ea02da132a0d4c07182c","content":"| 实际输入改变 | 问句 | 保存的输出 |\n| --- | --- | --- |\n| 原图左红、右蓝 | 蓝杯左边的杯子 | 左红杯 |\n| 灰块遮住左杯 | 同上 | 无唯一可见目标 |\n| 只交换两杯 RGB 色值 | 同上 | 无目标，蓝杯已在最左 |\n| 保持交换图，只把 left 改 right | 蓝杯右边的杯子 | 右红杯 |","sha256":"0de466cf5672ac86c9dd4214b39ae872dd46ee32d4256b3291c661a7a495c361","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"grounding-object-contract","title":"一句话怎样落实为正确对象的区域？"},"page":{"id":"reading-opening","title":"已保存的指代实验：从像素提框后再应用关系","number":1},"sectionTitle":"已保存的指代实验：从像素提框后再应用关系","currentQuestion":"已保存的指代实验：从像素提框后再应用关系","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"table":{"headers":["实际输入改变","问句","保存的输出"],"rows":[["实际输入改变","问句","保存的输出"],["原图左红、右蓝","蓝杯左边的杯子","左红杯"],["灰块遮住左杯","同上","无唯一可见目标"],["只交换两杯 RGB 色值","同上","无目标，蓝杯已在最左"],["保持交换图，只把 left 改 right","蓝杯右边的杯子","右红杯"]]},"previousBlock":"b-efb11cb39a4df88a2ec5a2b4","nextBlock":"b-37ddfcb7277dc75a119b9b58","prerequisites":[{"id":"b-97d3bd79b5a7dadf94c8a8e4","url":"#b-97d3bd79b5a7dadf94c8a8e4","role":"section-title"},{"id":"b-bd15a843c17a310fed9ce2ad","url":"#b-bd15a843c17a310fed9ce2ad","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-37ddfcb7277dc75a119b9b58","type":"p","scope":"grounding-pixel-interventions","url":"https://codingai-lec03.pages.dev/course/blocks/grounding-pixel-interventions.html#b-37ddfcb7277dc75a119b9b58","content":"遮挡是真的修改图片，不是在答案字典里删除目标；交换色值保留轮廓和背景；语言对照保持图片完全相同。代码断言干预区域外像素不变，并保存差分图。它因此能够支持这些有限条件下“图片怎样改变候选、关系怎样选择目标”的解释。","sha256":"55fb5f3349686e772ea73d62b88bc243f3fdb3efd45df8c2454262f799c64329","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"grounding-object-contract","title":"一句话怎样落实为正确对象的区域？"},"page":{"id":"reading-opening","title":"已保存的指代实验：从像素提框后再应用关系","number":1},"sectionTitle":"已保存的指代实验：从像素提框后再应用关系","currentQuestion":"已保存的指代实验：从像素提框后再应用关系","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-32a3ea02da132a0d4c07182c","nextBlock":"b-175046355e8512b5674f783a","prerequisites":[{"id":"b-97d3bd79b5a7dadf94c8a8e4","url":"#b-97d3bd79b5a7dadf94c8a8e4","role":"section-title"},{"id":"b-bd15a843c17a310fed9ce2ad","url":"#b-bd15a843c17a310fed9ce2ad","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-175046355e8512b5674f783a","type":"p","scope":"grounding-pixel-interventions","url":"https://codingai-lec03.pages.dev/course/blocks/grounding-pixel-interventions.html#b-175046355e8512b5674f783a","content":"![输入干预与返回框](/assets/display/969fa088b392b7fb8a846fb3.webp)","sha256":"70aab124a4de4ccb1629c768a54fc062777fd23a1d521943a55fc0d99ea993f7","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"grounding-object-contract","title":"一句话怎样落实为正确对象的区域？"},"page":{"id":"reading-opening","title":"已保存的指代实验：从像素提框后再应用关系","number":1},"sectionTitle":"已保存的指代实验：从像素提框后再应用关系","currentQuestion":"已保存的指代实验：从像素提框后再应用关系","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-37ddfcb7277dc75a119b9b58","nextBlock":"b-01c0b1a977cb2249943ab387","prerequisites":[{"id":"b-97d3bd79b5a7dadf94c8a8e4","url":"#b-97d3bd79b5a7dadf94c8a8e4","role":"section-title"},{"id":"b-bd15a843c17a310fed9ce2ad","url":"#b-bd15a843c17a310fed9ce2ad","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-01c0b1a977cb2249943ab387","type":"p","scope":"grounding-pixel-interventions","url":"https://codingai-lec03.pages.dev/course/blocks/grounding-pixel-interventions.html#b-01c0b1a977cb2249943ab387","content":"旧关键词规则错选蓝杯仍保留为失败基线。当前结果来自 2026-09-22 已保存运行，本轮未重跑。若以后换自然照片，颜色阈值和杯子类别约定会失效；那时才需要加载相应视觉模型，并对输入和实际框做新的可追溯评价。","sha256":"2c856e57e1c37c3d70093e2c5de7c9890f3a1bfe364eef8129588edb5dd7a622","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"grounding-object-contract","title":"一句话怎样落实为正确对象的区域？"},"page":{"id":"reading-opening","title":"已保存的指代实验：从像素提框后再应用关系","number":1},"sectionTitle":"已保存的指代实验：从像素提框后再应用关系","currentQuestion":"已保存的指代实验：从像素提框后再应用关系","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-175046355e8512b5674f783a","nextBlock":"b-aec2bf693ad82a4e7e46cca0","prerequisites":[{"id":"b-97d3bd79b5a7dadf94c8a8e4","url":"#b-97d3bd79b5a7dadf94c8a8e4","role":"section-title"},{"id":"b-bd15a843c17a310fed9ce2ad","url":"#b-bd15a843c17a310fed9ce2ad","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-aec2bf693ad82a4e7e46cca0","type":"h2","scope":"grounding-pixel-interventions","url":"https://codingai-lec03.pages.dev/course/blocks/grounding-pixel-interventions.html#b-aec2bf693ad82a4e7e46cca0","content":"## 继续阅读与实验","sha256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"grounding-object-contract","title":"一句话怎样落实为正确对象的区域？"},"page":{"id":"reading-opening","title":"已保存的指代实验：从像素提框后再应用关系","number":1},"sectionTitle":"已保存的指代实验：从像素提框后再应用关系","currentQuestion":"已保存的指代实验：从像素提框后再应用关系","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-01c0b1a977cb2249943ab387","nextBlock":null,"prerequisites":[{"id":"b-97d3bd79b5a7dadf94c8a8e4","url":"#b-97d3bd79b5a7dadf94c8a8e4","role":"section-title"},{"id":"b-bd15a843c17a310fed9ce2ad","url":"#b-bd15a843c17a310fed9ce2ad","role":"problem-context"}],"relatedBlocks":[]}]}