Hierarchical video object segmentation

In this paper, we propose a general video object segmentation framework which views object segmentation from a unified Bayesian perspective and optimizes the MAP formulated problem in a progressive manner. Based on object detection and tracking results, a three-level hierarchical video object segmentation approach is presented. At the first level, an offline learned segmentor is applied to each object tracking result of current frame to get a coarse segmentation. At the second level, the coarse segmentation is updated into an intermediate segmentation by a temporal model which propagates the fine segmentation of previous frame to current frame based on a discriminative feature points voting process. At the third level, the intermediate segmentation is refined by an iterative procedure which uses online collected color-and-shape information to get the final result. We apply the approach to pedestrian segmentation on many challenging datasets that demonstrates its effectiveness.