Removing Mistracking of Multibody Motion Video Database Hopkins155

Many mathematical techniques have been presented for classifying feature point trajectories over multibody motion video sequences into different motions, and most are applied to the Hopkins155 database for evaluating their performance. In this paper, we point out that Hopkins155 has problems and that it cannot necessarily evaluate the performance correctly. We create a new database by removing incorrect trajectories from Hopkins155. The basic principle of mistracking removal lies on the fact that correct trajectories all belong to parallel 2-D affine spaces in a high-dimensional space if all motions are translational and that parallel 2-D affine spaces are included in a 3-D affine space. Noting that if the image sequence is divided into short intervals, individual motions can be regarded as approximately translational in each interval, we detect incorrect trajectories by repeated plane fitting in the 3-D space using RANSAC. We point out why conventional outlier removal procedure does not work and demonstrate that out method allows us to tell in which frames incorrect trajectories occurred. The performance of multibody motion segmentation can be correctly evaluated using our database.