Unverified50% confidenceCautionExact time
原始边界框坐标需转换为模型专用的 token 表示，任何格式偏差都可能导致模型无法正确学习定位能力
1
Sources
50%
Confidence
Long-term
Relevance
7/11/2026
First Seen
Sources
微调PaliGemma 2实现自定义目标检测：从通用到专用的实践指南
redditr/learnmachinelearning7/10/2026
Related Claims
Unverified自回归生成的天然缺陷在于模型生成第一个 token 时无法看到整个输出序列的全局结构，导致早期决策难以被后续内容纠正70% similarUnverified纯坐标/模板匹配方案的脆弱性在于版式变化零容忍，字体大小、行间距、列顺序的细微调整都会导致解析失败69% similarUnverified标准自回归模型逐Token从左到右生成文本，在时间步t仅依赖之前的输出，缺乏自我审查和回溯机制，导致错误积累效应66% similarUnverifiedIf an occluding object cuts through the outer ring of a SIMITTAG marker, marker detection will fail cleanly, because the pose depends entirely on the ellipse fitting of the outer ring contour63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/486021API
curl https://kongchang.com/api/v1/knowledge/claims/486021MCP
get_claim(id=486021)