Dynamic Dense Graph Convolutional Network for Skeleton-Based Human Motion Prediction.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Graph Convolutional Networks (GCNs) are successful in skeleton-based human motion prediction via neural message passing.
• Challenges remain in optimal graph construction and message passing for GCNs in this domain.

Purpose of the Study:

• To address limitations in GCN graph construction and message passing for human motion prediction.
• To introduce a novel Dynamic Dense Graph Convolutional Network (DD-GCN) for enhanced performance.

Main Methods:

• Developed a Dynamic Dense Graph Convolutional Network (DD-GCN) model.
• Constructed a dense graph using 4D adjacency modeling for comprehensive motion representation.
• Implemented an integrated dynamic message passing framework for sample-specific relevance learning.

Main Results:

• DD-GCN significantly outperforms state-of-the-art GCN-based methods on benchmark datasets (Human 3.6M, CMU Mocap).
• The model shows particular effectiveness in long-term and extremely long-term human motion prediction protocols.

Conclusions:

• The proposed DD-GCN effectively solves key challenges in GCN-based human motion prediction.
• Dynamic dense graph construction and message passing offer superior performance, especially for extended motion sequences.