Foreground soft segmentation for search space reduction

This paper proposes the foreground soft segmentation to reduce the search space for human pose estimation. Our contributions are twofold: (i) using a quad-tree structure to efficiently obtain a binary label image for foreground; and (ii) search space reduction by limiting the locations within the only area based on segmented foreground. We show that the proposed method achieves segmentation of foreground in the various environments and poses.