관절 좌표 추정을 위한 다중 학습 방법

Abstract

In the field of human pose estimation, lack of dataset has always been a solid problem. Many research tried to overcome this issue by applying additional information such as depth information, body part segmentation, etc. In this paper, we focus on the effects of various information which are keypoint, silhouette and body part segmentation. By analyzing each cases, we propose an optimized network for human pose estimation.