Robust Labeling and Invariance Modeling for Unsupervised Cross-Resolution Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Cross-resolution person re-identification (CR-ReID) matches individuals across low-resolution (LR) and high-resolution (HR) images.
• Existing unsupervised CR-ReID methods often use computationally expensive cross-resolution fusion for pseudo-labels and features.

Purpose of the Study:

• To propose an efficient unsupervised CR-ReID framework using a single encoder.
• To enhance the robustness and accuracy of CR-ReID models.

Main Methods:

• Developed a Robust Labeling and Invariance Modeling (RLIM) framework with a single encoder.
• Introduced Cross-Resolution Robust Labeling (CRL) for accurate pseudo-label generation.
• Implemented Random Texture Augmentation (TexA) to improve robustness against noisy textures.
• Utilized resolution-cluster consistency loss for learning resolution-invariant features.

Main Results:

• RLIM framework significantly outperforms existing unsupervised CR-ReID methods.
• Achieved performance comparable to some supervised CR-ReID approaches.
• Demonstrated effectiveness across multiple benchmark datasets.

Conclusions:

• The proposed RLIM framework offers an efficient and effective solution for unsupervised CR-ReID.
• The method successfully addresses the challenges of resolution gaps and noisy data.
• RLIM shows strong potential for real-world person re-identification applications.