Decoupled High-Dimensional Spatial Pose-Block for 3D Human Pose Estimation

This paper investigates the problem of estimating 3D human pose from monocular image, one of the most formidable tasks in computer vision due to depth ambiguity, occlusion and intricate hierarchical arrangement of human body parts. Our proposed method consists of several decoupled sub-networks that demonstrate the effectiveness of decoupled architecture and high-dimensional projection on features of neural network. Contrary to the curse of dimensionality, we show that deploying high-dimensional linear projection on input and other features supervises neurons to learn spatial reasoning of images more accurately. Our 3D pose estimator takes 2D pose predictions and directly learns to predict poses in 3D space. We also propose a spatial pose-block that considers the complicated hierarchy and structure of human body and further refines the predicted pose from 3D predictor for final output. We validate our 3D pose estimation model on Human 3.6M benchmark under three protocols. Empirical evaluations show that our approach achieves performance competitive to state-of-the-art on similar benchmark.