Multi-feature balanced network for clothes-changing person re-identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Clothes-changing person re-identification (CC-ReID) is crucial for long-term surveillance.
• Extracting clothing-irrelevant features is a major challenge due to significant clothing variations.
• Existing methods often fail to fully utilize all available information in pedestrian images.

Purpose of the Study:

• To propose a novel Multi-features Balanced Network (MBNet) for robust CC-ReID.
• To enhance the extraction of clothing-unrelated features for improved re-identification accuracy.
• To overcome limitations of current methods that do not fully leverage pedestrian image data.

Main Methods:

• Developed a Multi-features Balanced Network (MBNet) with three branches: global, clothing-unrelated, and mask.
• Introduced a knowledge transfer module (KTM) to highlight clothing-unrelated clues.
• Incorporated a feature attention module (FAM) and a cross fusion module (CFM) to refine feature extraction and integration.

Main Results:

• Achieved competitive Rank-1/mAP accuracies of 44.6%/22.7% on one dataset.
• Demonstrated strong performance with 58.3%/57.9% and 87.2%/84.0% on two other datasets.
• The proposed MBNet significantly improves robustness against clothing changes.

Conclusions:

• MBNet effectively exploits clothing-unrelated features for superior CC-ReID performance.
• The integration of multiple branches and specialized modules enhances feature discriminability and contextual information.
• The approach shows superiority on both synthetic and realistic datasets, validating its effectiveness.