Object segmentation from sparse views of wide-baseline images

In this paper, we propose an automatic approach to segment object from multiple sparse views of wide-baseline images. Depth and occlusion information recovered from multiple views assist the object initialization and segmentation processes. The initial object patch is extract based on a saliency map incorporating the depth and locality cues. We then formulate the object segmentation task as an energy minimization problem, which is solved by graph cut optimization. Based on the basic energy function, local background modeling, adaptive data fusion and 3D graph construction are developed to make the segmentation toward better results. Experimental results on self-recorded images and the benchmarks demonstrate the efficiency and robustness of the proposed approach.