Robust human body segmentation based on part appearance and spatial constraint

Human body segmentation in images is desirable in various practical applications, e.g., content-based image retrieval. However, it remains a challenging problem due to various body poses and confusing background. To overcome these difficulties, two properties of human body are explored in this paper, i.e., complementary property and weak structure property. Complementary property means that different human body parts always have the similar appearances. With this property, we propose to construct the Part Appearance Map (PAM). PAM can effectively represent the appearance probability of what a pixel belong to a human body, even for inaccurate human pose obtained by pictorial structure model. Afterward, robust foreground and background seeds are acquired by PAM. To utilize the structure information of human body effectively, we propose a novel graph cuts method - spatial constraint based graph cuts (SCGC), which incorporates weak structure property of human body parts into the cost function. The weak structure property constrains the arms, legs and head to appear in limited space under the condition that the location of torso is ascertained. With this property, the SCGC can successfully remove false segmentations by traditional graph cuts methods due to their similar appearances to human body. Experimental results show that the proposed method achieves promising performance and outperforms many state-of-the-art methods over publicly available challenging datasets which contain arbitrary poses.