知识卡片
像素损失会偏向大区域
逐像素交叉熵使占面积大的类别贡献更多梯度,小目标可能被忽视。按频率或面积加权能调整关注分配,但也可能放大稀有噪声,权重需结合数据与任务验证;类别频率与单实例大小也不是同一件事。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 目标识别
原书出处 · 1 处
- · 查看原书依据
This loss might focus too much on the large objects and ignore small objects. Therefore, it is useful to add a weight to each class that normalizes each class according to its average area in the training set.
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:摘录中关于像素损失会偏向大区域的定义、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:This loss might focus too much on the large objects and ignore small objects. Therefore, it is useful to add a weight to each class that normalizes each class according to its average area in the training set.