A Comparison between Anatomy-Based and Data-Driven Tree Models for Human Pose Estimation

Abstract

Tree structures are commonly used to model relationships between body parts for articulated Human Pose Estimation (HPE). Tree structures can be used to model relationships among feature maps of joints in a structured learning framework using Convolutional Neural Networks (CNNs). This paper proposes new data-driven tree models for HPE. The data-driven tree structures were obtained using the Chow-Liu Recursive Grouping (CLRG) algorithm, representing the joint distribution of human body joints and tested using the Leeds Sports Pose (LSP) dataset. The paper analyzes the effect of the variation of the number of nodes on the accuracy of the HPE. Experimental results showed that the data-driven tree model obtained 1% higher HPE accuracy compared to the traditional anatomy-based model. A further improvement of 0.5% was obtained by optimizing the number of nodes in the traditional anatomy-based model.