Learning a Quality-Based Ranking for Feature Point Trajectories

Long term motion analysis poses many standing challenges that need to be addressed for advancing this field. One of these challenges is finding algorithms that correctly handle occlusion and can detect when a pixel trajectory needs to be stopped. Very few optical algorithms provide an occlusion map and are appropriate for this task. Another challenge is finding a framework for the accurate evaluation of the motion field produced by an algorithm. This work makes two contributions in these directions. First, it presents a RMSE based error measure for evaluating feature tracking algorithms on sequences with rigid motion under the affine camera model. The proposed measure was observed to be consistent with the relative ranking of a number of optical flow algorithms on the Middlebury dataset. Second, it introduces a feature tracking algorithm based on RankBoost that automatically prunes bad trajectories obtained by an optical flow algorithm. The proposed feature tracking algorithm is observed to outperform many feature trackers based on optical flow using both the proposed measure and an indirect measure based on motion segmentation.