A Real-Time Memory Updating Strategy for Unsupervised Person Re-Identification

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised person re-identification (ReID) heavily relies on clustering-based methods.
• Memory-based contrastive learning is effective for unsupervised representation learning but suffers from inaccurate cluster proxies and momentum updating strategies.

Purpose of the Study:

• To address limitations in current unsupervised person ReID methods.
• To propose a novel real-time memory updating strategy (RTMem) to enhance contrastive learning for ReID.

Main Methods:

• Introduced RTMem, a strategy that updates cluster centroids with current mini-batch features without momentum.
• Developed two contrastive losses: sample-to-instance loss for dataset-wide relationships and sample-to-cluster loss for instance-to-cluster alignment.
• Utilized pseudo-labels from a density-based clustering algorithm for sample-to-cluster loss.

Main Results:

• RTMem enables up-to-date features for each cluster, overcoming momentum-based limitations.
• The proposed method achieved a 9.3% performance improvement on the Market-1501 dataset compared to the baseline.
• Outperformed state-of-the-art unsupervised person ReID methods across three benchmark datasets.

Conclusions:

• RTMem is an effective strategy for unsupervised person ReID, enhancing contrastive learning.
• The proposed approach offers significant performance gains and outperforms existing methods.
• The methodology provides a robust solution for unsupervised representation learning in ReID tasks.