Anatomy-Aware 3D Human Pose Estimation With Bone-Based Pose Decomposition

In this work, we propose a new solution to 3D human pose estimation in videos. Instead of directly regressing the joint locations, draw inspiration from skeleton anatomy and decompose task into bone direction prediction length prediction, which locations can be completely derived. Our motivation is fact that lengths remain consistent across time. This promotes us develop effective techniques utilize global information all frames video for high-accuracy prediction. Moreover, network, fully-convolutional propagating architecture with long skip connections. Essentially, it predicts directions different bones hierarchically without using any time-consuming memory units (e.g. LSTM). A novel shift loss further introduced bridge training networks. Finally, employ an implicit attention mechanism feed 2D keypoint visibility scores model as extra guidance, significantly mitigates depth ambiguity many challenging poses. full outperforms previous best results on Human3.6M MPI-INF-3DHP datasets, where comprehensive evaluation validates effectiveness our model.