SteadyFlow: Spatially Smooth Optical Flow for Video Stabilization

我们通过光流初始化staedyFlow，然后通过时空分析丢弃不连续运动，并通过运动补偿填充缺失区域。

通过时空分析不连续的光流区域

（1）：空间分析：
章节4.2：In the spatial domain, we threshold the gradient magnitude of raw optical flow to identify discontinuous regions. Once the magnitude at p is larger than the threshold (0.1 in our experiment), p is considered as ’outlier’。

检查原始光流的梯度大小，当梯度>0.1时，则认为p是外点。这种办法只能检测到移动目标的边界。因此，我们需要时间分析来定位目标
（2）时间分析：
首先，定义ct§ = ∑t ut§,其中p表示图像的某个像素pixel，ut§表示像素p在时间为t的帧的位移，ct§表示从0到t时间内的位移矢量和。如果没有运动的区域，位移矢量和应该是连续的。在发生运动的区域，位移矢量和是快速变化的。
所以，通过定义Mt来获得外点p.

通过运动补偿填充缺失区域

参考:
[14] Content-preserving warps for 3d video stabilization
[15] Subspace video stabilization
[17] Video stabilization with a depth camera

[14]

迭代优化的步骤：

At each iteration, the first step is to exclude outliers and fill in the missing regions of the input SteadyFlow according to the mask Mt. The motion completion is described in Section 4.3. The second step is to stabilize the SteadyFlow, which will be described in Section 5. In the third step, the stabilized SteadyFlow is then used to further refine Mt by temporal analysis of discontinuous flow vectors as described in Section 4.2.