Tracking by Deblatting

Abstract Objects moving at high speed along complex trajectories often appear in videos, especially videos of sports. Such objects travel a considerable distance during exposure time single frame, and therefore, their position the frame is not well defined. They as semi-transparent streaks due to motion blur cannot be reliably tracked by general trackers. We propose novel approach called Tracking Deblatting based on observation that directly related intra-frame trajectory an object. Blur estimated solving two intertwined inverse problems, blind deblurring image matting, which we call deblatting. By postprocessing, non-causal estimates continuous, complete, accurate object for whole sequence. Tracked are precisely localized with higher temporal resolution than conventional Energy minimization dynamic programming used detect abrupt changes motion, bounces. High-order polynomials then fitted smooth segments between The output continuous function assigns location every real-valued stamp from zero number frames. proposed algorithm was evaluated newly created dataset high-speed camera using Trajectory-IoU metric generalizes traditional Intersection over Union measures accuracy trajectory. method outperforms baselines both recall accuracy. Additionally, show precise physical calculations possible, such radius, gravity, sub-frame velocity. Velocity estimation compared measurements radars. Results performance terms Trajectory-IoU, recall, velocity estimation.