← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8891 · 生成模型:GPT-6

采样高斯不保精确卷积闭包

理解等级:高级

把连续高斯逐点采样再截断,通常不能保持连续族的精确卷积规律。连续σ参数也未必等于归一化离散核的实际空间标准差。级联比较应计算实际离散矩、支撑和响应,不能仅把标称σ平方相加就声称输出逐值相等。

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

      You can check that if you compute the approximation for σ2 = 1 by discretizing the Gaussian, the result obtained is not equal to doing g5 ◦ g5. Therefore, as you apply successive convolutions of discretized Gaussians the errors will accumulate. That is, the convolution of discretized Gaussians is not a Gaussian anymore.

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第17章《模糊滤波器》(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/c017b.xhtml) - 结论依据:摘录中关于采样高斯不保精确卷积闭包的定义、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:You can check that if you compute the approximation for σ2 = 1 by discretizing the Gaussian, the result obtained is not equal to doing g5 ◦ g5. Therefore, as you apply successive convolutions of discretized Gaussians the errors will accumulate. That is, the convolution of discretized Gaussians is not a Gaussian anymore.