知识卡片
检测位置应随图像变换
对象平移或缩放后,其类别应尽可能保持,输出框则应相应平移缩放,即位置等变而非位置数值不变。枚举窗口和尺度可扩大搜索覆盖,却不保证离散尺度、边界及裁剪下实现严格的连续不变性。
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
In this approach, location and translation invariance are achieved by the bounding box proposal mechanism. We can add another invariance, such as rotation invariance, by proposing rotated bounding boxes. 50.4.1.2 Selective search The window scanning approach can be slow as the same classifier needs to be applied to tens of thousands of image patches. Selective search makes the process more efficient by proposing an initial
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:摘录中关于检测位置应随图像变换的定义、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:In this approach, location and translation invariance are achieved by the bounding box proposal mechanism. We can add another invariance, such as rotation invariance, by proposing rotated bounding boxes. 50.4.1.2 Selective search The window scanning approach can be slow as the same classifier needs to be applied to tens of thousands of image patches. Selective search makes the process more efficient by proposing an initial