Human pose estimation using consistent max-covering

We propose a novel consistent max-covering scheme for human pose estimation. Consistent max-covering formulates pose estimation as the covering of body part polygons on an object silhouette so that the body part tiles maximally cover the foreground, match local image features, and satisfy body linkage plan and color constraints. It uses high order constraints to anchor multiple body parts simultaneously; the hyper-edges in the part relation graph are essential for detecting complex poses. Because of using multiple clues in pose estimation, this method is resistant to cluttered foregrounds. We propose an efficient linear relaxation method to solve the consistent max-covering problem. Experiments on a variety of images and videos show that the proposed method is more robust than locally constrained methods for human pose estimation.