This paper presents a player segmentation approach based on 3D model hypotheses for soccer games. We use a hyperplane model for player modeling and a collection of piecewise geometric models for background modeling. To determine the assignment of each pixel in the image plane, we test it with two model hypotheses. We construct a cost function that measures the fitness of model hypotheses for each pixel. To fully utilize the perspective diversity of the multiview imagery, we propose a three-step strategy to choose the best model for each pixel. The experimental results show that our segmentation approach based on 3D model hypotheses outperforms conventional temporal median and graph cut methods for both subjective and objective evaluation.
QC 20130111