cvpr07-benchmark-final

A Benchmark for the Comparison of 3-D Motion Segmentation AlgorithmsRoberto Tron ReneV´ idalCenter for Imaging Science, Johns Hopkins University308B Clark Hall, 3400 N. Charles St., Baltimore MD 21218, USAhttp://www.vision.jhu.eduAbstract have been developed (see§2 for a brief review). How-ever, all existing techniques have been typically evalu-Over the past few years, several methods for segment- ated on a handful of sequences, with limited compari-ing a scene containing multiple rigidly moving objects have son against other methods. This motivates a study onbeen proposed. However, most existing methods have been the real performances of these methods.tested on a handful of sequences only, and each method has2. Perspective methods assume a perspective projectionbeen often tested on a different set of sequences. Therefore,model. In this case, point trajectories associated withthe comparison of different methods has been fairly limited.each moving object lie in a multilinear variety (bilinearIn this paper, we compare four 3-D motion segmentation al-for two views, trilinear for three views, etc.) There-gorithms for affine cameras on a benchmark of 155 motionfore, motion segmentation is equivalent to clusteringsequences of checkerboard, traffic, and articulated scenes.these multilinear varieties. Because this problem isnontrivial, most prior work has been limited to alge-1. Introduction braic methods for factorizing bilinear and trilinear va-rieties (see e ...