Deeply Learned View-Invariant Features for Cross-View Action Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human action classification from diverse viewpoints presents significant challenges due to substantial data variations.
• Developing discriminative features that are invariant to different views is crucial for robust action recognition.

Purpose of the Study:

• To propose novel deep learning models for learning view-specific and view-shared features for human action classification.
• To enhance the robustness of features against view variations and improve classification accuracy.

Main Methods:

• Utilized deep models to learn both view-specific features (capturing unique dynamics) and view-shared features (encoding common patterns).
• Introduced a novel sample-affinity matrix to balance information transfer within and across multiple views for shared feature learning.
• Encouraged incoherence between feature types to reduce redundancy and promoted geometric closeness for features within the same categories.

Main Results:

• The proposed sample-affinity matrix effectively learns discriminative shared features robust to view variations.
• Learned features demonstrate improved discriminative power by encouraging intra-class feature points to be geometrically closer.
• Stacked feature layers result in robust view-invariant features.

Conclusions:

• The developed approach significantly outperforms state-of-the-art methods on three multi-view datasets for human action classification.
• The method effectively addresses the challenge of view variations in action recognition by learning robust, view-invariant features.