GAReID: Grouped and Attentive High-Order Representation Learning for Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Person re-identification (ReID) faces challenges due to misaligned human body parts in detected bounding boxes.
• Existing methods often struggle with part misalignment, impacting recognition accuracy.

Purpose of the Study:

• To propose an effective framework, Grouped Attentive Re-Identification (GAReID), for learning part-aligned and background-robust representations for person ReID.
• To address the issue of part misalignment without relying on landmark detection or explicit feature partitioning.

Main Methods:

• The GAReID framework incorporates Grouped High-Order Pooling (GHOP) and Attentive High-Order Pooling (AHOP) layers.
• These layers generate high-order image and foreground features, respectively.
• A novel Grouped Kronecker Product (GKP) is introduced for efficient high-order feature compression using channel grouping and shuffling.

Main Results:

• The proposed method effectively reduces part misalignments.
• GAReID demonstrates superior representational capabilities through compressed high-order features.
• The framework achieves state-of-the-art performance across multiple person ReID benchmarks.

Conclusions:

• The GAReID framework offers an interpretable and effective solution for person re-identification.
• It successfully handles part misalignment and background robustness, outperforming existing methods.
• The theoretical and experimental results validate the superiority of the GAReID approach for person ReID.