Abstract

Occlusion degrades the performance of human pose estimation. In this paper,we introduce targeted keypoint and body part occlusion attacks. The effects ofthe attacks are systematically analyzed on the best performing methods. Inaddition, we propose occlusion specific data augmentation techniques againstkeypoint and part attacks. Our extensive experiments show that human poseestimation methods are not robust to occlusion and data augmentation does notsolve the occlusion problems.

 

Quick Read (beta)