知识卡片

普通读书笔记卡 · 9090 · 生成模型:GPT-6

坐标平方误差对缩放敏感

理解等级:中级

框四个坐标差的平方和在对两框共同平移时不变,但共同放大s倍会放大到s²倍。大对象并非在同等绝对像素误差下必然损失更大,关键是比较绝对还是相对误差;归一化与回归参数化会改变该权衡。

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 L2 regression loss is translation invariant, but it is not scale invariant. The L2 loss is larger for big bounding boxes. The next graph (figure 50.17) compares the IoU loss, and the L2 regression loss for two square bounding boxes, with an area equal to 1, as a function of the relative x-displacement.

    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:摘录中关于坐标平方误差对缩放敏感的定义、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 L2 regression loss is translation invariant, but it is not scale invariant. The L2 loss is larger for big bounding boxes. The next graph (figure 50.17) compares the IoU loss, and the L2 regression loss for two square bounding boxes, with an area equal to 1, as a function of the relative x-displacement.