3D Human Pose Regression Using Graph Convolutional Network

Abstract

3D human pose estimation is a difficult task, due to challenges such as occluded body parts and ambiguous poses. Graph convolutional networks encode the structural information of the human skeleton in the form of an adjacency matrix, which is beneficial for better pose prediction. We propose one such graph convolutional network named PoseG-raphNet for 3D human pose regression from 2D poses. Our network uses an adaptive adjacency matrix and kernels specific to neighbor groups. We evaluate our model on the Hu-man3.6M dataset which is a standard dataset for 3D pose estimation. Our model's performance is close to the state-of-theart, but with much fewer parameters. The model learns interesting adjacency relations between joints that have no physical connections, but are behaviorally similar.

MoCoSys: Human Motion Correction based on Deep Learning Coupled with 3D+t Laplacian Motion Representation

“…The idea to integrate temporal information to improve the spatial accuracy was then explored. The belief that temporal information can provide a solution to the lack of depth information has led to new approaches using several consecutive images or a sequence of 2D joint positions [1,12,[21][22][23][24][25][26][27]. These approaches use sequence-to-pose or sequence-to-sequence neural networks, such as recurrent, temporal convolution, or spatio-temporal graph convolution neural networks to work with the sequence of 2D data (images or 2D joint locations).…”

On the Role of Depth Predictions for 3d Human Pose Estimation

“…To account for these relationships, some methods incorporate "static" anthropometric constraints and regularization procedures, while others are based on temporal architectures that infer these dependencies across video frames [31,2,16,11]. For instance, people have used graph convolutional and graph attention networks to naturally model spatial relationships between joints while building lightweight networks [75,76,78,79].…”