{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/blocks/glip-language-objective.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"GLIP：把类别和短语都放到对象级定位任务中","url":"https://codingai-lec03.pages.dev/course/blocks/glip-language-objective.html","markdown":"https://codingai-lec03.pages.dev/course/blocks/glip-language-objective.md","pages":[{"id":"reading-opening","title":"GLIP：把类别和短语都放到对象级定位任务中","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-9a1ebc25b5c8e62f006e4ebc","targetId":"b-9a1ebc25b5c8e62f006e4ebc","page":"reading-opening","operation":"move","beforeSHA256":"7442d3ce01ebb10319c1251fb001b79a97a34d33c62e26ac22717e7e636af05b","afterSHA256":"7442d3ce01ebb10319c1251fb001b79a97a34d33c62e26ac22717e7e636af05b","reason":"Same source body in authored semantic page"},{"oldId":"b-7f2314ec50499f0215657baf","targetId":"b-7f2314ec50499f0215657baf","page":"reading-opening","operation":"move","beforeSHA256":"7d8bc1408f3f30ddf3278f32dbb54c475cf47f3b9b5ab00d4d7f1ef8f718627d","afterSHA256":"7d8bc1408f3f30ddf3278f32dbb54c475cf47f3b9b5ab00d4d7f1ef8f718627d","reason":"Same source body in authored semantic page"},{"oldId":"b-f679b3c8911d029edfcad7e7","targetId":"b-f679b3c8911d029edfcad7e7","page":"reading-opening","operation":"move","beforeSHA256":"24baff1cf87b2de7ca97ba1336e8c33d4560ab39283b2ef8c7c1c36dba4b50fc","afterSHA256":"24baff1cf87b2de7ca97ba1336e8c33d4560ab39283b2ef8c7c1c36dba4b50fc","reason":"Same source body in authored semantic page"},{"oldId":"b-2af7f622d2ee5d28db9a5673","targetId":"b-2af7f622d2ee5d28db9a5673","page":"reading-opening","operation":"move","beforeSHA256":"80dfe0ea15830ea4065fc5c1744a90f46c42c9fa95ba56ae9d2715668a597d60","afterSHA256":"80dfe0ea15830ea4065fc5c1744a90f46c42c9fa95ba56ae9d2715668a597d60","reason":"Same source body in authored semantic page"},{"oldId":"b-b68ef36a5ed53c0b115031b4","targetId":"b-b68ef36a5ed53c0b115031b4","page":"reading-opening","operation":"move","beforeSHA256":"bb4028ef924f2a8d235dbf0f1ba9b2b386877b47814234f5a6e6b5a0f191e153","afterSHA256":"bb4028ef924f2a8d235dbf0f1ba9b2b386877b47814234f5a6e6b5a0f191e153","reason":"Same source body in authored semantic page"},{"oldId":"b-0743be96d3a232c1fe0f33af","targetId":"b-0743be96d3a232c1fe0f33af","page":"reading-opening","operation":"move","beforeSHA256":"acb849912de164fda2c6b1f36086f33e94a4c1f8fc5b5957a303c674c13a844f","afterSHA256":"acb849912de164fda2c6b1f36086f33e94a4c1f8fc5b5957a303c674c13a844f","reason":"Same source body in authored semantic page"},{"oldId":"b-d235e24fb08d752a6a08c1d9","targetId":"b-d235e24fb08d752a6a08c1d9","page":"reading-opening","operation":"rewrite","beforeSHA256":"256ea3d770d2bfec6f3e634211424333d834313d39f9d45ee30c2a46686427b3","afterSHA256":"2418b5a849472004c12de1ce96e54ae079c6b3964eeadbca4a10743281606293","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-c1b226e9ef9ab95852b961df","targetId":"b-c1b226e9ef9ab95852b961df","page":"reading-opening","operation":"move","beforeSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","afterSHA256":"2b9ccef7e1220e2250cceff1abd8bdd736b42f8da5a223f26b029f44b962e807","reason":"Same source body in authored semantic page"}],"blocks":[{"id":"b-9a1ebc25b5c8e62f006e4ebc","type":"h1","scope":"glip-language-objective","url":"https://codingai-lec03.pages.dev/course/blocks/glip-language-objective.html#b-9a1ebc25b5c8e62f006e4ebc","content":"# GLIP：把类别和短语都放到对象级定位任务中","sha256":"7442d3ce01ebb10319c1251fb001b79a97a34d33c62e26ac22717e7e636af05b","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":"GLIP：把类别和短语都放到对象级定位任务中","number":1},"sectionTitle":"GLIP：把类别和短语都放到对象级定位任务中","currentQuestion":"GLIP：把类别和短语都放到对象级定位任务中","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-7f2314ec50499f0215657baf","prerequisites":[{"id":"b-7f2314ec50499f0215657baf","url":"#b-7f2314ec50499f0215657baf","role":"problem-context"}],"relatedBlocks":[{"id":"b-d235e24fb08d752a6a08c1d9","url":"#b-d235e24fb08d752a6a08c1d9","relation":"figure-and-explanation"}]},{"id":"b-7f2314ec50499f0215657baf","type":"p","scope":"glip-language-objective","url":"https://codingai-lec03.pages.dev/course/blocks/glip-language-objective.html#b-7f2314ec50499f0215657baf","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":"GLIP：把类别和短语都放到对象级定位任务中","number":1},"sectionTitle":"GLIP：把类别和短语都放到对象级定位任务中","currentQuestion":"GLIP：把类别和短语都放到对象级定位任务中","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-9a1ebc25b5c8e62f006e4ebc","nextBlock":"b-f679b3c8911d029edfcad7e7","prerequisites":[{"id":"b-9a1ebc25b5c8e62f006e4ebc","url":"#b-9a1ebc25b5c8e62f006e4ebc","role":"section-title"}],"relatedBlocks":[{"id":"b-d235e24fb08d752a6a08c1d9","url":"#b-d235e24fb08d752a6a08c1d9","relation":"figure-and-explanation"}]},{"id":"b-f679b3c8911d029edfcad7e7","type":"p","scope":"glip-language-objective","url":"https://codingai-lec03.pages.dev/course/blocks/glip-language-objective.html#b-f679b3c8911d029edfcad7e7","content":"GLIP 将对象检测与短语定位用于共同的预训练。传统检测类别可以转写成文字概念，短语资料又提供更开放的语言—区域对应。视觉与文字发生交互，训练同时要求语义匹配和位置正确；这比给一组外部裁图计算 CLIP 整图分数多了任务结构与定位监督。","sha256":"24baff1cf87b2de7ca97ba1336e8c33d4560ab39283b2ef8c7c1c36dba4b50fc","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":"GLIP：把类别和短语都放到对象级定位任务中","number":1},"sectionTitle":"GLIP：把类别和短语都放到对象级定位任务中","currentQuestion":"GLIP：把类别和短语都放到对象级定位任务中","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7f2314ec50499f0215657baf","nextBlock":"b-2af7f622d2ee5d28db9a5673","prerequisites":[{"id":"b-9a1ebc25b5c8e62f006e4ebc","url":"#b-9a1ebc25b5c8e62f006e4ebc","role":"section-title"},{"id":"b-7f2314ec50499f0215657baf","url":"#b-7f2314ec50499f0215657baf","role":"problem-context"}],"relatedBlocks":[{"id":"b-d235e24fb08d752a6a08c1d9","url":"#b-d235e24fb08d752a6a08c1d9","relation":"figure-and-explanation"}]},{"id":"b-2af7f622d2ee5d28db9a5673","type":"p","scope":"glip-language-objective","url":"https://codingai-lec03.pages.dev/course/blocks/glip-language-objective.html#b-2af7f622d2ee5d28db9a5673","content":"原论文使用 2700 万 grounding 训练样本，包含人工标注和由网络图文扩展的资料；在其协议中，直接评估 COCO 得到 49.8 AP，LVIS 得到 26.9 AP，且说明预训练没有使用 COCO 图像。这里的数字是论文所报告的对象级迁移结果，不是本课程持球图的运行。它支持语言条件和区域学习能够带来可迁移检测接口这一方向。","sha256":"80dfe0ea15830ea4065fc5c1744a90f46c42c9fa95ba56ae9d2715668a597d60","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":"GLIP：把类别和短语都放到对象级定位任务中","number":1},"sectionTitle":"GLIP：把类别和短语都放到对象级定位任务中","currentQuestion":"GLIP：把类别和短语都放到对象级定位任务中","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f679b3c8911d029edfcad7e7","nextBlock":"b-b68ef36a5ed53c0b115031b4","prerequisites":[{"id":"b-9a1ebc25b5c8e62f006e4ebc","url":"#b-9a1ebc25b5c8e62f006e4ebc","role":"section-title"},{"id":"b-7f2314ec50499f0215657baf","url":"#b-7f2314ec50499f0215657baf","role":"problem-context"}],"relatedBlocks":[{"id":"b-d235e24fb08d752a6a08c1d9","url":"#b-d235e24fb08d752a6a08c1d9","relation":"figure-and-explanation"}]},{"id":"b-b68ef36a5ed53c0b115031b4","type":"p","scope":"glip-language-objective","url":"https://codingai-lec03.pages.dev/course/blocks/glip-language-objective.html#b-b68ef36a5ed53c0b115031b4","content":"对课堂场景，“拿球的蓝衣人”要求图中多个对象、属性与关系共同约束一个框。若候选框身份错了，后接分割模型即使边界非常准确，也会分错对象。反过来，目标身份正确但边界粗糙，应检查分割与采样接口。把定位和掩码衔接时分开检查，正是第二部分仍然有用的原因。","sha256":"bb4028ef924f2a8d235dbf0f1ba9b2b386877b47814234f5a6e6b5a0f191e153","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":"GLIP：把类别和短语都放到对象级定位任务中","number":1},"sectionTitle":"GLIP：把类别和短语都放到对象级定位任务中","currentQuestion":"GLIP：把类别和短语都放到对象级定位任务中","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-2af7f622d2ee5d28db9a5673","nextBlock":"b-0743be96d3a232c1fe0f33af","prerequisites":[{"id":"b-9a1ebc25b5c8e62f006e4ebc","url":"#b-9a1ebc25b5c8e62f006e4ebc","role":"section-title"},{"id":"b-7f2314ec50499f0215657baf","url":"#b-7f2314ec50499f0215657baf","role":"problem-context"}],"relatedBlocks":[{"id":"b-d235e24fb08d752a6a08c1d9","url":"#b-d235e24fb08d752a6a08c1d9","relation":"figure-and-explanation"}]},{"id":"b-0743be96d3a232c1fe0f33af","type":"p","scope":"glip-language-objective","url":"https://codingai-lec03.pages.dev/course/blocks/glip-language-objective.html#b-0743be96d3a232c1fe0f33af","content":"LAB15 负责透明的语言关系规则，GLIP 论文负责更一般的学习方案与研究证据。两者互相解释，不互相冒充。主讲保留这一对象级接口；完整训练配置与更广数据对照留材料阅读。","sha256":"acb849912de164fda2c6b1f36086f33e94a4c1f8fc5b5957a303c674c13a844f","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":"GLIP：把类别和短语都放到对象级定位任务中","number":1},"sectionTitle":"GLIP：把类别和短语都放到对象级定位任务中","currentQuestion":"GLIP：把类别和短语都放到对象级定位任务中","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b68ef36a5ed53c0b115031b4","nextBlock":"b-d235e24fb08d752a6a08c1d9","prerequisites":[{"id":"b-9a1ebc25b5c8e62f006e4ebc","url":"#b-9a1ebc25b5c8e62f006e4ebc","role":"section-title"},{"id":"b-7f2314ec50499f0215657baf","url":"#b-7f2314ec50499f0215657baf","role":"problem-context"}],"relatedBlocks":[{"id":"b-d235e24fb08d752a6a08c1d9","url":"#b-d235e24fb08d752a6a08c1d9","relation":"figure-and-explanation"}]},{"id":"b-d235e24fb08d752a6a08c1d9","type":"figure","scope":"glip-language-objective","url":"https://codingai-lec03.pages.dev/course/blocks/glip-language-objective.html#b-d235e24fb08d752a6a08c1d9","content":"图内标签与数值：指代：拿球的蓝衣人；“蓝衣人”；“拿球的”；图文交互 + 区域预测；短语对应与定位监督；训练目标落在局部区域；输出：与短语对应的框；GLIP 把词语与空间定位放进同一学习任务；教学框展示任务接口，不是 GLIP 运行结果；区域交互和定位学习不能由整图分数直接代替。\n\n短语最终要对应局部区域；网页框是课程标注，不是本轮 GLIP 输出。 查看对应图解状态\n\n- 指代：拿球的蓝衣人\n- “蓝衣人”\n- “拿球的”\n- 图文交互 + 区域预测\n- 短语对应与定位监督\n- 训练目标落在局部区域\n- 输出：与短语对应的框\n- GLIP 把词语与空间定位放进同一学习任务\n- 教学框展示任务接口，不是 GLIP 运行结果；区域交互和定位学习不能由整图分数直接代替。短语最终要对应局部区域；网页框是课程标注，不是本轮 GLIP 输出。 [查看对应图解状态](/#/l3-r2-p63/0)[查看完整图解](/#/l3-r2-p63/0)","sha256":"2418b5a849472004c12de1ce96e54ae079c6b3964eeadbca4a10743281606293","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":"GLIP：把类别和短语都放到对象级定位任务中","number":1},"sectionTitle":"GLIP：把类别和短语都放到对象级定位任务中","currentQuestion":"GLIP：把类别和短语都放到对象级定位任务中","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"短语最终要对应局部区域；网页框是课程标注，不是本轮 GLIP 输出。 查看对应图解状态","labels":["指代：拿球的蓝衣人","“蓝衣人”","“拿球的”","图文交互 + 区域预测","短语对应与定位监督","训练目标落在局部区域","输出：与短语对应的框","GLIP 把词语与空间定位放进同一学习任务","教学框展示任务接口，不是 GLIP 运行结果；区域交互和定位学习不能由整图分数直接代替。"],"description":"短语最终要对应局部区域；网页框是课程标注，不是本轮 GLIP 输出。 查看对应图解状态指代：拿球的蓝衣人“蓝衣人”“拿球的”图文交互 + 区域预测短语对应与定位监督训练目标落在局部区域输出：与短语对应的框GLIP 把词语与空间定位放进同一学习任务教学框展示任务接口，不是 GLIP 运行结果；区域交互和定位学习不能由整图分数直接代替。","evidenceIdentity":"短语最终要对应局部区域；网页框是课程标注，不是本轮 GLIP 输出。 查看对应图解状态","images":[{"src":"/assets/display/ab33341ce4021f383054cfc5.webp","alt":""}]},"previousBlock":"b-0743be96d3a232c1fe0f33af","nextBlock":"b-c1b226e9ef9ab95852b961df","prerequisites":[{"id":"b-9a1ebc25b5c8e62f006e4ebc","url":"#b-9a1ebc25b5c8e62f006e4ebc","role":"section-title"},{"id":"b-7f2314ec50499f0215657baf","url":"#b-7f2314ec50499f0215657baf","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-c1b226e9ef9ab95852b961df","type":"h2","scope":"glip-language-objective","url":"https://codingai-lec03.pages.dev/course/blocks/glip-language-objective.html#b-c1b226e9ef9ab95852b961df","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":"GLIP：把类别和短语都放到对象级定位任务中","number":1},"sectionTitle":"GLIP：把类别和短语都放到对象级定位任务中","currentQuestion":"GLIP：把类别和短语都放到对象级定位任务中","sourceSlides":[],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-d235e24fb08d752a6a08c1d9","nextBlock":null,"prerequisites":[{"id":"b-9a1ebc25b5c8e62f006e4ebc","url":"#b-9a1ebc25b5c8e62f006e4ebc","role":"section-title"},{"id":"b-7f2314ec50499f0215657baf","url":"#b-7f2314ec50499f0215657baf","role":"problem-context"}],"relatedBlocks":[{"id":"b-d235e24fb08d752a6a08c1d9","url":"#b-d235e24fb08d752a6a08c1d9","relation":"figure-and-explanation"}]}]}