Arbitrary body segmentation in static images

In this paper, a novel method for segmenting arbitrary human body in static images is proposed. With the body probability map obtained by the pictorial structure model, we develop a superpixel based EM-like algorithm to refine the map, which can then serve as the seeds of graph cuts optimization. To better obtain the final segmentation, we propose a novel ℓ1ℓ1 based graph cuts algorithm, which uses the sparse coding to construct the initialized graph and calculates the terminal links (t-links) and neighborhood links (n-links  ) simultaneously from the constructed graph. By employing this ℓ1ℓ1 based graph cuts, we can effectively and efficiently segment the human body from static images. The experiments on the publicly available challenging datasets demonstrate that our method outperforms many state-of-the-art methods on human body segmentation.

► We propose a powerful framework to segment human body from static images.
► An efficient superpixel based EM-like algorithm is developed.
► Both top-down pictorial structure and bottom-up superpixel are considered.
► A novel ℓ1ℓ1 based graph cuts algorithm is proposed.