Arbitrary Body Segmentation With a Novel Graph Cuts-Based Algorithm

Abstract

This letter presents a novel superpixel based graph cuts algorithm for human body segmentation from images. The standard graph cuts algorithm constructs the t-links and n-links separately. While in this letter, we argue that the construction of these two links can be build simultaneously. The proposed algorithm starts from the construction of a complete similarity graph based on superpixels where the t-links and n-links have been embedded and hence the t-links and n-links can be easily obtained using the max pooling function and distance matrix respectively. This strategy not only makes the segmentation more accurate but also makes the method more robust to the selection of parameters. The experiments on two challenging public datasets validate that our method can segment the object more accurately than the standard graph cuts, Grabcut and geodesic star convexity graph cuts with a few user provided seeds and is very robust to the parameters changes.