Particle Video Revisited: Tracking Through Occlusions Using Point Trajectories

Tracking pixels in videos is typically studied as an optical flow estimation problem, where every pixel described with a displacement vector that locates it the next frame. Even though wider temporal context freely available, prior efforts to take this into account have yielded only small gains over 2-frame methods. In paper, we revisit Sand and Teller’s “particle video” approach, study tracking long-range motion trajectory multiple future frames. We re-build classic approach using components drive current state-of-the-art object tracking, such dense cost maps, iterative optimization, learned appearance updates. train our models amodal point trajectories mined from existing data synthetically augment multi-frame occlusions. test benchmarks keypoint label propagation tasks, compare favorably against feature