Automatic Pose Quality Assessment for Adaptive Human Pose Refinement

Abstract

Multi-person pose estimation from a 2D image is an essential technique for many computer vision tasks. Although the development of deep convolutional neural networks has brought large improvement to human pose estimation, some complex cases are still challenging to even state-of-the-art approaches. The forms of people in the images are diverse. The quality of estimated poses is difficult to guarantee. Estimated poses usually cannot be directly used in practical application scenarios. In this paper, we propose a pose quality assessment model and an adaptive human pose refinement method. The pose quality assessment model can measure per-joint pose quality with a quality score and select qualified estimated poses. The adaptive pose refinement method can handle each estimated pose respectively, until reaching a certain standard. Our experiments show the effectiveness of the pose quality assessment model and confirm that adaptive pose refinement method performs better than generally refining all poses once. Our adaptive pose refinement method reaches state-of-the-art performance.