Disentangled Sample Guidance Learning for Unsupervised Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised person re-identification (Re-ID) faces challenges due to the absence of ground truth labels.
• Existing methods often rely on iterative clustering for pseudo-label generation, struggling with sample selection for effective learning.

Purpose of the Study:

• To propose a novel Disentangled Sample Guidance Learning (DSGL) method to enhance unsupervised Re-ID.
• To address the critical problem of selecting high-confidence and discriminative samples for training.

Main Methods:

• DSGL comprises Disentangled Sample Mining (DSM) to separate identity-relevant and irrelevant factors in images.
• DSM constructs disentangled positive/negative groups for discriminative information extraction.
• Discriminative Feature Learning (DFL) integrates these groups using specialized loss functions and regularization for improved person distinctiveness.

Main Results:

• DSGL significantly boosts mAP by 6.6% (ResNet50) and 0.6% (ViT) on Market-1501.
• mAP improvements of 10.1% (ResNet50) and 6.9% (ViT) were observed on MSMT17.
• DSGL outperforms state-of-the-art methods on Market-1501, MSMT17, PersonX, and VeRi-776 datasets.

Conclusions:

• The proposed DSGL method effectively improves unsupervised person Re-ID performance.
• DSGL's approach to sample selection and feature learning enhances the model's ability to distinguish individuals.
• The method demonstrates superior results across multiple benchmark datasets, validating its effectiveness.