EXAM: A Framework of Learning Extreme and Moderate Embeddings for Person Re-ID

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (Re-ID) faces challenges from variations in human pose, bounding box alignment, and complex backgrounds.
• Existing Re-ID methods struggle to effectively capture both salient and common human features simultaneously.

Purpose of the Study:

• To propose a new framework, EXAM (Extreme And Moderate feature embeddings), for enhancing person re-identification performance.
• To develop a method that learns discriminative features using attention-based guidance.

Main Methods:

• EXAM utilizes discriminative feature learning with attention-based guidance during training.
• Extreme embeddings capture salient human features via global max-pooling.
• Moderate embeddings capture common human features via average-pooling.
• Joint supervision is achieved using multiple triplet and cross-entropy loss functions.

Main Results:

• The framework integrates attention deduction with discriminative feature learning in an end-to-end manner.
• Comparative experiments and ablation studies demonstrate the effectiveness of the proposed EXAM framework.
• The learned feature representation by EXAM achieves state-of-the-art performance in Re-ID tasks.

Conclusions:

• The EXAM framework offers an effective solution for person re-identification.
• The proposed approach successfully learns discriminative features by leveraging both extreme and moderate embeddings.
• EXAM's attention-guided mechanism contributes to achieving top-tier performance in Re-ID.