Pose and Shape Estimation of Occluded Humans with Attention and Conditional GAN

The occlusions of body parts often appear in the images, which makes the human pose and shape estimation from single images difficult. This paper proposes a novel single-image oriented framework to tackle this problem, where two effective tactics are proposed. One is a multi-scale attention module which generates the enhanced multi-scale attention features with rich contextual information, so that efficient global pose and shape distribu-tion can be obtained without the affection of occlusion. The other is heatmap based conditional generative ad-versarial networks (GAN) which utilize the poses from the joint heatmaps as constraints and thus can refine the mesh of the occluded subject accurately. Combining these two tactics can make the proposed human pose and shape estimation method robustly recover the body meshes with both global prediction and local details. Qualita-tive and quantitative experiments show the efficiency of the proposed method for occluded humans.