Intra-Inter Domain Similarity for Unsupervised Person Re-Identification

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised person Re-Identification (ReID) methods often struggle with domain discrepancy across cameras, leading to inaccurate pseudo-label generation.
• Existing approaches fail to adequately account for variations in feature similarity across different camera domains.

Purpose of the Study:

• To develop a robust unsupervised person ReID method that mitigates domain discrepancy.
• To improve the accuracy of pseudo-label computation in multi-camera environments.

Main Methods:

• Decomposed similarity computation into intra-domain and inter-domain stages.
• Utilized CNN features for intra-domain similarity and classification scores for inter-domain similarity.
• Proposed Instance and Camera Style Normalization (ICSN) with TNorm to enhance robustness to domain variations.

Main Results:

• Achieved competitive performance on multiple datasets in unsupervised, intra-camera supervised, and domain generalization settings.
• Demonstrated significant improvement over recent unsupervised methods, achieving 64.4% rank-1 accuracy on MSMT17.
• The proposed method effectively alleviates domain discrepancy, leading to more reliable pseudo-labels.

Conclusions:

• The proposed two-stage similarity computation and ICSN effectively address domain discrepancy in unsupervised person ReID.
• The method offers a significant advancement in ReID accuracy and robustness across diverse camera setups.
• This work provides a more reliable solution for person Re-Identification in real-world, multi-camera scenarios.