Unsupervised Person Re-Identification by Deep Asymmetric Metric Embedding

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Supervised person re-identification (Re-ID) models require extensive labeled cross-view data, limiting scalability.
• Unsupervised Re-ID models struggle with view-specific biases caused by variations in illumination, viewpoint, and occlusion.

Purpose of the Study:

• To develop a novel unsupervised deep learning framework for person re-identification.
• To address the challenge of view-specific bias in unsupervised Re-ID scenarios.
• To improve the accuracy and scalability of person re-identification across disjoint camera views.

Main Methods:

• Proposed an unsupervised asymmetric distance metric learning approach based on cross-view clustering.
• Developed a novel unsupervised loss function to integrate the asymmetric metric into a deep neural network.
• Introduced the DEep Clustering-based Asymmetric MEtric Learning (DECAMEL) framework, jointly learning feature representations and the asymmetric metric.

Main Results:

• DECAMEL effectively alleviates view-specific bias by learning view-specific feature transformations.
• The framework learns a compact cross-view cluster structure, facilitating the discovery of discriminative cross-view information.
• Experiments on seven benchmark datasets demonstrated the framework's effectiveness across varying data scales.

Conclusions:

• DECAMEL offers a scalable and effective solution for unsupervised person re-identification.
• The proposed asymmetric metric learning approach successfully tackles feature distortions caused by cross-view variances.
• This work advances unsupervised Re-ID by enabling robust identity matching without reliance on labeled data.