Euclidean-Distance-Preserved Feature Reduction for efficient person re-identification.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Person Re-identification (Re-ID) uses deep neural networks for matching person images across cameras.
• Current Re-ID methods create high-dimensional features, increasing computational and storage complexity.
• Existing feature reduction techniques have limitations in end-to-end optimization or theoretical guarantees.

Purpose of the Study:

• To propose a novel method, Euclidean-Distance-Preserving Feature Reduction (EDPFR), to address the complexity of high-dimensional features in Re-ID.
• To combine the strengths of reduction-after-training and reduction-during-training methods for improved Re-ID.
• To introduce a feature-level distillation loss for more flexible and efficient knowledge transfer.

Main Methods:

• EDPFR formulates feature reduction as matrix decomposition with a condition to preserve Euclidean distances.
• The matrix decomposition is integrated into deep neural networks for end-to-end optimization and batch training.
• A novel feature-level distillation loss is proposed, leveraging the Euclidean-Distance-Preserving property for direct feature space distillation.

Main Results:

• EDPFR effectively reduces feature dimensions while preserving Euclidean distances between features (L2(fa,fb)=L2(fa',fb')).
• The proposed method achieves improved flexibility and efficiency in knowledge distillation by performing it directly in the feature space.
• Experiments on Market-1501, DukeMTMC-reID, and MSMT datasets demonstrate the effectiveness of EDPFR.

Conclusions:

• EDPFR offers a theoretically sound and practically efficient solution for reducing feature dimensions in person re-identification.
• The integration of EDPFR into deep networks enables robust and optimized Re-ID models.
• The novel feature-level distillation enhances the applicability and performance of knowledge distillation in Re-ID tasks.