Human pose estimation based on lightweight basicblock
Motion-region annotation for complex videos via label propagation across occluders