Person Reidentification via Discrepancy Matrix and Matrix Metric

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Person reidentification (re-id) is crucial for video surveillance and forensics.
• Existing methods often rely on appearance-based feature vectors or labeled distance metrics.
• Human cognition offers insights for improved re-id strategies.

Purpose of the Study:

• To propose a new pattern for person reidentification based on human cognitive processes.
• To transform feature descriptions from characteristic vectors to discrepancy matrices.
• To develop a novel matrix-based distance metric for enhanced identification.

Main Methods:

• Feature description is converted from a characteristic vector to a discrepancy matrix.
• A matrix metric, comprising intradiscrepancy and interdiscrepancy projection parts, is employed.
• An objective function with consistent and discriminative terms is optimized using gradient descent and alternating optimization.

Main Results:

• The proposed discrepancy matrix pattern effectively represents person features.
• The matrix metric demonstrates superior performance compared to traditional vector metrics.
• Experimental results on public datasets validate the approach's effectiveness.

Conclusions:

• The novel discrepancy matrix pattern offers a more robust approach to person reidentification.
• This method improves identification accuracy in surveillance and forensic contexts.
• The findings suggest a promising new direction for re-id research.