知识卡片

普通读书笔记卡 · 9077.c · 生成模型:GPT-6

迭代配准重新线性化残差

理解等级:高级

以当前流变形第二帧,再求残余增量并更新,可避免只在零位移做一次泰勒展开。一般梯度下降一步不等于Lucas–Kanade求解的正规方程一步;Gauss–Newton式更新与普通梯度下降有不同预条件,迭代次数也不保证精度单调改善。

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

      We can now run gradient descent on this loss. Running only one iteration would be similar to the gradient-based optical flow algorithm described in the previous section. However, adding iterations will provide a more accurate estimate of optical flow. At each iteration n, the estimated optical flow will be used to compute the warped frame ℓ1(x + un, y + vn, t + 1) and

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第48章《光流估计》(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/c048.xhtml) - 结论依据:摘录中关于迭代配准重新线性化残差的定义、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:We can now run gradient descent on this loss. Running only one iteration would be similar to the gradient-based optical flow algorithm described in the previous section. However, adding iterations will provide a more accurate estimate of optical flow. At each iteration n, the estimated optical flow will be used to compute the warped frame ℓ1(x + un, y + vn, t + 1) and