Foreground object extraction from multiview images with layer quantization and boundary refinement

In this paper, we present a method of extracting a foreground object from multiview images with layer quantization and boundary refinement. The method has two steps, quantization of the disparity map with PSO (Particle Swarm Optimization) algorithm and refinement step. The disparity map estimated by a simple window-based method is quantized into three layers, and the initial foreground mask is created with it. After detecting suspicious regions with errors in boundaries of the initial foreground mask, they are refined using the statistical measure, the Bhattacharyya distance. The proposed method is not dependent on specific disparity estimation methods, and the experimental results show that it extracts a foreground object from multiview images accurately and robustly.