← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 9081.a · 生成模型:GPT-6

训练损失须处理掩码和尺度

理解等级:中级

真实标签可能只覆盖部分像素,损失应在有效位置计算并明确归一化分母。不同分辨率的流以对应像素单位表示,直接比较前需换算;平方和与均值也会随图像大小产生不同权重,影响训练与结果解释。

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 sum is over all the pixels in the image. Each pixel has an estimated optical flow . 49.2.1.3 Handling occlusions One of the challenges for estimating optical flow is that, as objects move, they will occlude some pixels from the background and reveal new ones. Therefore, together with the estimated flow it is also convenient to detect occluded pixels. If we have ground truth

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第49章《学习运动估计》(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/c049.xhtml) - 结论依据:摘录中关于训练损失须处理掩码和尺度的定义、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 sum is over all the pixels in the image. Each pixel has an estimated optical flow . 49.2.1.3 Handling occlusions One of the challenges for estimating optical flow is that, as objects move, they will occlude some pixels from the background and reveal new ones. Therefore, together with the estimated flow it is also convenient to detect occluded pixels. If we have ground truth