Multi-Biometric Unified Network for Cloth-Changing Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Pattern Recognition

Background:

• Person re-identification (re-ID) typically struggles with identifying individuals who change clothing.
• Clothing variations significantly impact the performance of existing re-ID models, as attire occupies large pixel areas and can mislead identification.

Purpose of the Study:

• To develop a robust person re-identification model capable of handling cloth-changing scenarios.
• To exploit clothing-independent cues for improved re-ID accuracy.

Main Methods:

• Proposed a novel Multi-biometric Unified Network (MBUNet) incorporating a multi-biological feature branch (head, neck, shoulders) to resist cloth changes.
• Integrated a differential feature attention module (DFAM) for discriminative biological feature extraction.
• Employed a differential recombination on max pooling (DRMP) strategy and a direction-adaptive graph convolutional layer for robust global and pose features.
• Introduced a Lightweight Domain Adaptation Module (LDAM) for enhancing transferable features across scenarios.
• Incorporated mAP optimization into the objective function for joint training.

Main Results:

• The MBUNet demonstrated significant advantages in cloth-changing re-ID tasks.
• Extensive experiments on five datasets validated the model's effectiveness.

Conclusions:

• The proposed MBUNet offers a robust solution for person re-identification in challenging cloth-changing scenarios.
• Exploiting clothing-independent biological features is crucial for improving re-ID performance when attire varies.