Efficient motion capture data recovery via relationship-aggregated graph network and temporal pattern reasoning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Computer Graphics

Background:

• Human motion capture (mocap) data is vital for realistic character animation.
• Missing optical marker data due to occlusions or detachment hinders real-world mocap applications.
• Existing mocap data recovery methods face challenges with articulated complexity and long-term movement dependencies.

Purpose of the Study:

• To propose an efficient and effective approach for recovering missing human motion capture data.
• To address the limitations of current methods in handling complex articulated movements and temporal dependencies.
• To enhance the performance and reliability of mocap data in animation and related fields.

Main Methods:

• Developed a Relationship-aggregated Graph Network (RGN) with Local Graph Encoder (LGE) and Global Graph Encoder (GGE) to encode skeletal structure and relationships.
• Implemented Temporal Pattern Reasoning (TPR) using self-attention and temporal transformers to capture intra-frame and long-term dependencies.
• Integrated LGE, GGE, and TPR into the RGN-TPR framework for comprehensive spatio-temporal feature extraction.

Main Results:

• The RGN-TPR framework demonstrated superior performance in mocap data recovery compared to state-of-the-art methods.
• Extensive experiments on public datasets validated the qualitative and quantitative effectiveness of the proposed approach.
• The method successfully recovered missing motion data, improving the accuracy and completeness of mocap sequences.

Conclusions:

• The RGN-TPR framework offers an efficient and robust solution for human motion capture data recovery.
• The proposed method effectively addresses the challenges of articulated complexity and long-term dependencies in motion data.
• This work contributes to advancing realistic character animation and other applications reliant on accurate mocap data.