Clothing-invariant contrastive learning for unsupervised person re-identification

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Clothing change person re-identification (CC-ReID) is crucial for matching individuals across images with varying attire.
• Existing CC-ReID methods predominantly rely on supervised learning, demanding extensive manual annotation.
• The need for efficient, unsupervised CC-ReID solutions is significant.

Purpose of the Study:

• To develop a novel unsupervised framework for CC-ReID, addressing the limitations of supervised approaches.
• To enable accurate person re-identification despite significant clothing variations.
• To reduce the reliance on manually annotated datasets in CC-ReID tasks.

Main Methods:

• Proposed a clothing-invariant contrastive learning (CICL) framework for unsupervised CC-ReID.
• Introduced Random Clothing Augmentation (RCA) to generate clothing-change positive pairs efficiently.
• Developed Semantic Fusion Clustering (SFC) for unsupervised pseudo-label generation and Semantic Alignment Contrastive loss (SAC loss) for robust feature learning.

Main Results:

• The CICL framework demonstrated superior performance compared to existing unsupervised CC-ReID methods.
• Achieved competitive results, even rivaling supervised CC-ReID approaches on multiple datasets.
• The proposed RCA and SFC methods effectively facilitate clothing-invariant learning and identity-related feature extraction.

Conclusions:

• The CICL framework offers a powerful unsupervised solution for CC-ReID, overcoming the challenge of clothing variations.
• This approach significantly reduces the need for manual annotations, making CC-ReID more practical.
• The method shows strong potential for real-world applications requiring robust person re-identification across diverse scenarios.