← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 9089.c · 生成模型:GPT-6

IoU损失在不重叠时有平台

理解等级:中级

普通1−IoU在两框严格不相交的一片区域内恒为一,局部梯度不能告诉预测框应往哪移动。坐标回归能提供其他信号,但数值尺度不同;不能从“可写成损失”推断所有初始化都能仅靠该项优化。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      The IoU loss becomes 1 when the two bounding boxes do not overlap, which means that there will be no gradient information when running backpropagation. Now we can put together the classification and the localization loss to compute the overall loss: (50.17) where the indicator function 𝟙(yi = 0) sets to zero the location loss for the bounding boxes that do not overlap with any

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第50章《目标识别》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c050b.xhtml) - 结论依据:摘录中关于IoU损失在不重叠时有平台的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。 - 原始内容:The IoU loss becomes 1 when the two bounding boxes do not overlap, which means that there will be no gradient information when running backpropagation. Now we can put together the classification and the localization loss to compute the overall loss: (50.17) where the indicator function 𝟙(yi = 0) sets to zero the location loss for the bounding boxes that do not overlap with any