Learning a 3D Human Pose Distance Metric from Geometric Pose Descriptor.

Area of Science:

• Computer Vision
• Human Motion Analysis
• Machine Learning

Background:

• Estimating 3D pose similarity is crucial for analyzing 3D motion data.
• Existing methods using L2-like distances of joint coordinates or orientations inadequately capture human perceptual similarity.
• A need exists for more robust pose similarity metrics that align with human perception.

Purpose of the Study:

• To develop a novel pose distance metric for 3D motion data.
• To improve the accuracy and perceptual relevance of 3D pose similarity estimation.
• To create a foundational block for advanced 3D motion analysis applications.

Main Methods:

• Proposed a new feature set, Geometric Pose Descriptor (GPD), capturing geometric relations and temporal dynamics (velocities, accelerations).
• Developed a semi-supervised distance metric learning algorithm, Regularized Distance Metric Learning with Sparse Representation (RDSR), integrating unsupervised relationships and labels.
• Applied the metric to motion transition decision and content-based pose retrieval tasks.

Main Results:

• The Geometric Pose Descriptor (GPD) effectively encodes pose similarity.
• The Regularized Distance Metric Learning with Sparse Representation (RDSR) algorithm achieved superior results with limited human labels.
• Demonstrated improved performance in motion transition and pose retrieval applications.

Conclusions:

• The proposed pose distance metric, based on GPD and RDSR, offers a more perceptually relevant measure of 3D pose similarity.
• This method shows significant promise for various 3D motion-related applications, requiring minimal human supervision.
• The approach provides a strong foundation for future research in human motion analysis and understanding.