Multi-person Human Shape and Pose Estimation Based on Local and Global Optimization

Single-image based multi-person shape and pose estimation is challenging for crowded scenes. Current top-down methods often rely on pose priors which may cause error estimations due to inaccuracies. To address this limitation, this article presents a novel top-down multi-person mesh estimation framework via local and global combined optimization. Locally, random dropout and GAN are collaborated to generate initial shapes and poses, so that occlusions can be resisted. Globally, cross-attention between the centroid assisted initial poses and global image features is applied to optimize shapes and poses, considering the importance of global contexts in further overcoming the occlusion issues. Experiments on various benchmarks quantitatively and qualitatively confirm the efficacy of this method for crowded scenes.