Zhen Jianan ; Fang Qi ; Sun Jiaming ; Liu Wentao ; Jiang Wei ; Bao Hujun ; Zhou Xiaowei

Abstract

Recovering multi-person 3D poses with absolute scales from a single RGB imageis a challenging problem due to the inherent depth and scale ambiguity from asingle view. Addressing this ambiguity requires to aggregate various cues overthe entire image, such as body sizes, scene layouts, and inter-personrelationships. However, most previous methods adopt a top-down scheme thatfirst performs 2D pose detection and then regresses the 3D pose and scale foreach detected person individually, ignoring global contextual cues. In thispaper, we propose a novel system that first regresses a set of 2.5Drepresentations of body parts and then reconstructs the 3D absolute poses basedon these 2.5D representations with a depth-aware part association algorithm.Such a single-shot bottom-up scheme allows the system to better learn andreason about the inter-person depth relationship, improving both 3D and 2D poseestimation. The experiments demonstrate that the proposed approach achieves thestate-of-the-art performance on the CMU Panoptic and MuPoTS-3D datasets and isapplicable to in-the-wild videos.