Real-Time Selfie Video Stabilization

Figure 1. Our method stabilizes selfie videos using A background feature points and foreground face vertices in each frame. B The twobranch stabilization network infers C the moving least squares (MLS) warping for each frame. D We show the face and background motion of the input vs. our stabilized result. For visualization only, the background tracks are computed from the translation component of the homography between consecutive frames. The face tracks are computed from the centroid of the fitted face vertices in each frame.