Anchor Association Learning for Unsupervised Video Person Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video-based person re-identification (re-id) is crucial for video surveillance but faces challenges with supervised learning's reliance on extensive labeled data.
• Existing unsupervised re-id methods show unsatisfactory performance in real-world scenarios.

Purpose of the Study:

• To propose an effective unsupervised anchor association learning (UAAL) framework for video-based person re-identification.
• To overcome the limitations of supervised learning and improve the performance of unsupervised re-id methods.

Main Methods:

• The proposed UAAL framework treats feature representations of tracklets as anchors.
• Introduces an intracamera anchor association learning (IAAL) term for discriminative anchor learning within cameras.
• Employs an exponential moving average (EMA) strategy for anchor updates and an anchor memory module.
• Develops a cross-camera anchor association learning (CAAL) term using cyclic ranking and threshold filtering to identify positive anchor pairs across cameras.

Main Results:

• The UAAL framework achieves superior performance on public datasets.
• Achieved 73.2% rank-1 accuracy and 60.1% mean average precision (mAP) on the MARS dataset.
• Attained 89.7% rank-1 accuracy and 87.0% mAP on the DukeMTMC-VideoReID dataset.

Conclusions:

• The proposed unsupervised anchor association learning framework demonstrates significant advancements in video-based person re-identification.
• The method offers a viable solution for real-world surveillance applications where labeled data is scarce.