Learning Constrained Dynamic Correlations in Spatiotemporal Graphs for Motion Prediction.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Human motion prediction is complex due to spatiotemporal feature modeling challenges.
• Graph Convolutional Networks (GCNs) excel at modeling explicit connections in motion data.
• Current GCNs suffer from redundant parameters and inability to capture sample-specific motion variances.

Purpose of the Study:

• To develop a more efficient and accurate GCN for human motion prediction.
• To address parameter redundancy and sample-wise variance limitations in existing GCNs.
• To propose a novel dynamic spatiotemporal graph convolution (DSTD-GC) approach.

Main Methods:

• Introduced Dynamic Spatiotemporal Decompose Graph Convolution (DSTD-GC) with constrained dynamic correlation modeling.
• Parameterized common static constraints and dynamically extracted correspondence variances.
• Unified GCNs on spatiotemporal graphs and combined DSTD-GC with prior knowledge into DSTD-GCN.

Main Results:

• DSTD-GC uses 28.6% of the parameters compared to state-of-the-art GCNs.
• DSTD-GCN demonstrated superior performance on Human3.6M, CMU Mocap, and 3DPW datasets.
• Achieved 3.9%-8.7% higher prediction accuracy with 55.0%-96.9% fewer parameters.

Conclusions:

• DSTD-GC offers a more efficient and capable approach to human motion prediction.
• The dynamic modeling effectively captures sample-specific motion patterns.
• DSTD-GCN represents a significant advancement in human motion prediction accuracy and efficiency.