Model transfer from 2D to 3D study for boxing pose estimation.

Area of Science:

• Computer Vision
• Sports Science
• Machine Learning

Background:

• Boxing is gaining popularity on Chinese campuses, leading to a shortage of qualified coaches.
• Human pose estimation (HPE) technology offers a potential solution for training coaches and improving technique analysis.
• 3D cameras provide depth information, potentially enhancing HPE accuracy over 2D cameras, but face challenges with input channel consistency.

Purpose of the Study:

• To address the channel inconsistency issue between 2D and 3D imaging for HPE.
• To investigate the effectiveness of model transfer learning for boxing pose estimation.
• To analyze the performance of different network architectures (OpenPose, Hourglass, High-Resolution) in 3D HPE.

Main Methods:

• Implemented model transfer learning with channel patching to reconcile 2D and 3D input data.
• Utilized popular 2D HPE models: OpenPose, stacked Hourglass, and High-Resolution networks.
• Investigated reusing RGB channels to supplement depth information and analyzed key point differences.

Main Results:

• Model transfer learning improved average pose key point accuracies by 1-20% compared to baseline.
• 3D pose estimation achieved 0.3-0.5% higher accuracy than 2D methods.
• Stacked network structures outperformed parallel structures for specific key points (hip, knee), while parallel structures excelled on residual points.

Conclusions:

• Model transfer learning effectively bridges the gap between 2D and 3D data for boxing pose estimation.
• The study demonstrates a practical approach to enhance HPE for sports coaching applications.
• Network architecture choices impact performance on different body parts in 3D HPE.