Hand segmentation with metric learning superpixels

In this paper, we propose a novel method for hand segmentation. To improve the robustness to the wide range of hand appearances and illuminations, we segment the hand area with a superpixels based method instead of a general color model. With the exploitation of the distribution of hand pixels in color space, a distance metric learning stage is designed to promote the segmentation performance. This stage makes the points in hand areas more concentrate and pulls away from the points of background in color space. The comparisons with several widely used algorithms are made on both public available and our own datasets. The experimental results show the superior performance of our method.