Human body segmentation based on independent component analysis with reference at two-scale superpixel
Siyi Li, Huchuan Lu, Xiang Ruan, Yonglin Chen · IET Image Processing · 2012
In this study, a novel method to segment human body in static image is proposed based on independent component analysis with reference (ICA-R) at two-scale superpixel. In this work, the task is mainly decomposed into torso and lower body recovery. With the detected face, we obtain the reference signal of torso in the coarse torso region estimated by an augmented deformable torso model on the basis of the first-scale superpixel. The hip region is estimated based on the segmented torso for the lower body reference at the second-scale superpixel. Experiments on our dataset show that the proposed approach is robust and can accurately segment human body in images with a variety of poses, backgrounds and clothing.