Multi-Person Pose Regression With Distribution-Aware Single-Stage Models.

Area of Science:

• Computer Vision
• Machine Learning
• Human Pose Analysis

Background:

• Human posture understanding involves complex tasks like pose estimation, body mesh recovery, and pose tracking.
• Existing methods often require multiple passes or make simplifying assumptions about joint distributions, limiting accuracy and efficiency.

Purpose of the Study:

• To introduce a novel Distribution-Aware Single-stage (DAS) model for pose-related tasks.
• To enhance the accuracy and efficiency of single-pass human pose estimation, body mesh recovery, and pose tracking.

Main Methods:

• Developed a Distribution-Aware Single-stage (DAS) model that performs human position estimation and joint localization simultaneously in a single pass.
• Utilized normalizing flow to learn the true distribution of joint locations, providing a crucial prior for regression-based methods.
• Implemented a recursive update strategy to progressively refine regression targets, simplifying the regression process and improving performance.

Main Results:

• DAS achieves performance comparable to volumetric methods while offering a 1.5x speedup over previous best methods for multi-person pose estimation.
• Demonstrated state-of-the-art accuracy on CMU Panoptic and MuPoTS-3D datasets for multi-person pose estimation.
• Showcased effectiveness and efficiency for human body mesh recovery on 3DPW and pose tracking on PoseTrack2018 datasets.

Conclusions:

• The DAS model offers a superior, efficient, and accurate solution for various human pose-related tasks.
• The model's generality is proven across multi-person pose estimation, body mesh recovery, and pose tracking.
• The integration of normalizing flow and recursive updates significantly advances the field of human posture understanding.