Iterative Local-Global Collaboration Learning towards One-Shot Video Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video person re-identification (Re-ID) is crucial for surveillance but faces scalability issues with supervised methods requiring extensive labeled data.
• Unsupervised Re-ID methods are nascent, struggling with discriminative feature learning from unlabeled data.

Purpose of the Study:

• To develop a robust and scalable approach for one-shot video person re-identification.
• To improve the learning of discriminative person representations using unlabeled video data.

Main Methods:

• An iterative local-global collaboration learning framework combining global video and local frame sequence information.
• Incorporation of a variational information bottleneck for regularization to filter irrelevant features.
• A dynamic pseudo-label selection strategy to refine training data and improve model retraining.

Main Results:

• The proposed method effectively learns robust and discriminative person representations.
• Demonstrated superiority over several state-of-the-art competitors on public datasets (DukeMTMC-VideoReID and MARS).

Conclusions:

• The iterative local-global collaboration learning approach significantly advances one-shot video Re-ID.
• The method offers a promising solution for practical surveillance applications by overcoming data labeling limitations.