Distribution-Aware Single-Stage Models for Multi-Person 3D Pose Estimation

In this paper, we present a novel Distribution-Aware Single-stage (DAS) modelfor tackling the challenging multi-person 3D pose estimation problem. Differentfrom existing top-down and bottom-up methods, the proposed DAS modelsimultaneously localizes person positions and their corresponding body jointsin the 3D camera space in a one-pass manner. This leads to a simplifiedpipeline with enhanced efficiency. In addition, DAS learns the truedistribution of body joints for the regression of their positions, rather thanmaking a simple Laplacian or Gaussian assumption as previous works. Thisprovides valuable priors for model prediction and thus boosts theregression-based scheme to achieve competitive performance with volumetric-baseones. Moreover, DAS exploits a recursive update strategy for progressivelyapproaching to regression target, alleviating the optimization difficulty andfurther lifting the regression performance. DAS is implemented with a fullyConvolutional Neural Network and end-to-end learnable. Comprehensiveexperiments on benchmarks CMU Panoptic and MuPoTS-3D demonstrate the superiorefficiency of the proposed DAS model, specifically 1.5x speedup over previousbest model, and its stat-of-the-art accuracy for multi-person 3D poseestimation.