Progressive Direction-Aware Pose Grammar for Human Pose Estimation

Lu Zhou, Yingying Chen, Jinqiao Wang

Abstract: Human pose estimation is challenged by lots of factors such as complex articulation, occlusion and so on. Generally, message passing among human joints plays an important role in rectifying the wrong detection caused by referred challenges. In this paper, we propose a progressive direction-aware pose grammar model which performs message passing by building the pose grammar in a novel fashion. Firstly, a multi-scale Bi-C3D pose grammar module is proposed to promote message passing among human joints within a local range. We propose to conduct message passing by means of 3D convolution (C3D) which proves to be more effective compared with other sequential modeling techniques. To facilitate the message passing, we devise a novel adaptive direction guidance module where explicit direction information is embedded. Besides, we propose to fuse final results with attention maps to make full use of the bidirectional information and the fusion can be regarded as an ensemble process. Secondly, a more economic global regional grammar is introduced to build the relationships among human joints globally. The local-to-global modeling scheme promotes the message passing in a progressive manner and boosts the performance by a large margin. Promising results are achieved on MPII, LSP and COCO benchmarks.