Local Self-Expression Subspace Learning Network for Motion Capture Data.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Deep subspace learning is a key area of self-supervised learning.
• Existing methods often overlook the unique characteristics of temporal data and related tasks.

Purpose of the Study:

• To propose a novel local self-expression subspace learning network tailored for temporal data.
• To leverage motion capture data and segmentation tasks as supervision for enhanced feature learning.

Main Methods:

• Utilized temporal convolution modules for extracting temporal features from motion data.
• Introduced a local self-expression layer to maintain representation relationships with adjacent frames.
• Implemented group sparsity constraints and subspace projection loss to enhance feature representation and penalize clustering errors.

Main Results:

• The proposed model demonstrated superior performance on synthetic data segmentation.
• Achieved excellent results on three real-world motion capture data tasks.
• Validated the model's strong feature learning capabilities for temporal data.

Conclusions:

• The local self-expression subspace learning network effectively addresses limitations in current deep subspace learning for temporal data.
• The model's design enhances feature learning by considering temporality and local validity.
• The approach shows significant potential for applications in motion analysis and segmentation.