Discriminatively Unsupervised Learning Person Re-Identification via Considering Complicated Images

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised person re-identification (re-ID) methods typically cluster images and discard outliers.
• Outliers often represent complex real-world scenarios like occlusion, low resolution, and varied poses.
• Training solely on clustered data limits model robustness.

Purpose of the Study:

• To improve unsupervised person re-ID by incorporating complex, unclustered images.
• To develop a more robust feature extraction network capable of handling diverse image conditions.

Main Methods:

• Constructed a novel memory dictionary including both clustered and unclustered images.
• Designed a contrastive loss function that accounts for both image types.
• Trained a feature extraction network using the enhanced memory dictionary and loss function.

Main Results:

• The proposed memory dictionary and contrastive loss significantly improved person re-ID performance.
• Models demonstrated enhanced robustness in handling complicated images.

Conclusions:

• Incorporating unclustered, complex images is effective for unsupervised person re-ID.
• The developed method offers a more practical solution for real-world person re-identification challenges.