Spatio-temporal context for more accurate dense point trajectories estimation

Dense point trajectories estimation is a challenging yet important problem due to its potential of supporting other fields, such as motion estimation, action recognition, etc. In previous work, dense motion trackers always estimate trajectories based on consecutive frames and ignore scene context prior, thereby suffering from inaccurate estimation. In this paper, we present a novel dense point trajectories estimation framework which integrates trajectories spatio-temporal context into the estimation process. The spatial context for a trajectory refers to the support from its neighbouring trajectories, while the temporal context indicates the temporal appearance consistency for each trajectory. To obtain accurate and compact trajectories, we formulate the problem as an inference process in a Markov Random Field(MRF).We measure the accuracy of the algorithms on MIT sequences. Experimental results demonstrate that our methods can give more accurate dense point trajectories efficiently.