A weighted sparse coding model on product Grassmann manifold for video-based human gesture recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Classifying multi-dimensional data with complex geometry, like human gestures from videos, is challenging.
• Manifold structures effectively capture the intrinsic geometry of such data.
• Existing sparse coding on Grassmann manifolds loses spatial video information due to data vectorization.

Purpose of the Study:

• To represent human gesture videos on a product Grassmann manifold (PGM) while preserving spatial structure.
• To develop a weighted sparse coding model on PGM to account for feature importance.
• To propose an efficient optimization algorithm for learning coding coefficients.

Main Methods:

• Representing videos as data tensors and applying Higher Order Singular Value Decomposition (HOSVD) on the product Grassmann manifold (PGM).
• Developing a weighted sparse coding model on PGM, where weights signify the importance of factor manifolds (appearance, horizontal motion, vertical motion).
• Employing an optimization algorithm that embeds factor Grassmann manifolds into symmetric matrices space for learning coding coefficients.

Main Results:

• The proposed method effectively utilizes the spatial structure of videos through tensor representation.
• Experimental results on three public datasets demonstrate the method's competitiveness against existing approaches.
• The factor manifolds capture distinct features like appearance, horizontal, and vertical motion.

Conclusions:

• The weighted sparse coding on product Grassmann manifolds offers a robust approach for human gesture recognition.
• Preserving spatial video structure is crucial for improving classification accuracy.
• The method shows promising performance and potential for complex visual classification tasks.