Video object segmentation by clustering region trajectories

We propose a novel method to segment the moving object in video clips. In this work, we introduce a region trajectory generation model based on graph clustering. Point trajectories are widely used to measure the motion similarity because of their unambiguity. However, region trajectories preserve object boundaries, while optical flow based point trajectories always 'over-smooth' to the background. To cluster the region trajectories to meaningful objects, we employ a spectral embedding framework. Affinities are computed based on motion similarities between point trajectories associated with the region trajectories. Foreground topology is used in the discretization procedure to achieve robust segmentation, which is insensitive to the number of eigenvector selected. We validate our method on challenging dataset and provide statistical comparison with the state-of-the-art methods.