Cross-View Action Recognition Over Heterogeneous Feature Spaces.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Cross-view action recognition faces challenges due to differing data distributions and feature spaces between views.
• Transferring action models across heterogeneous features requires methods to bridge these differences.

Purpose of the Study:

• To develop a novel method for transferring action recognition models between source and target views with heterogeneous features.
• To create a discriminative common feature space for effective knowledge transfer.
• To address scenarios with limited or no labeled data in the target view.

Main Methods:

• Proposed heterogeneous transfer discriminant-analysis of canonical correlations (HTDCC) to learn a common feature space.
• Learned two projection matrices to map source and target data into the common space.
• Minimized inter-class canonical correlations and maximized intra-class canonical correlations while reducing data distribution mismatch.
• Introduced a weighting learning framework for adapting knowledge from multiple source views.

Main Results:

• HTDCC effectively discovers a discriminative common feature space for heterogeneous cross-view action recognition.
• The method demonstrates effectiveness even with few or no labeled samples in the target view.
• The weighting learning framework successfully leverages knowledge from multiple source views, assigning relevance-based weights.

Conclusions:

• HTDCC provides a robust solution for heterogeneous cross-view action recognition by learning a shared feature space.
• The proposed weighting framework enhances transfer learning by adaptively utilizing knowledge from multiple sources.
• The methods show significant promise for real-world applications requiring cross-view action understanding.