A GAN-Based Self-Training Framework for Unsupervised Domain Adaptive Person Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person re-identification (re-ID) is vital for surveillance and security, aiming to match pedestrians across non-overlapping camera views.
• Existing re-ID methods struggle with the lack of labeled data and domain shift between different camera systems.
• Generative adversarial networks (GANs) offer potential for unsupervised learning but require specialized frameworks for re-ID challenges.

Purpose of the Study:

• To develop a robust person re-identification (re-ID) framework capable of handling unlabeled target domain data and domain shift.
• To leverage generative adversarial networks (GANs) for effective feature extraction in challenging re-ID scenarios.
• To improve the accuracy and reliability of pedestrian identification across diverse and non-overlapping camera networks.

Main Methods:

• A self-training framework with progressive augmentation (SPA) was proposed, comprising a style transfer stage (STrans) and a self-training stage (STrain).
• The STrans stage utilized CycleGAN and Siamese Network for camera style transfer, preserving self-similarity and ensuring domain dissimilarity.
• The STrain stage employed alternating clustering and classification to progressively enhance model performance using both global and local features from target-domain images.

Main Results:

• The proposed SPA framework successfully addressed the challenges of unlabeled data and domain shift in person re-identification.
• The integrated style transfer and self-training approach enabled the extraction of robust features from target-domain data.
• The method achieved competitive accuracy compared to state-of-the-art approaches on established re-ID datasets.

Conclusions:

• The proposed GAN-based self-training framework with progressive augmentation offers a promising solution for person re-identification in real-world surveillance.
• The method effectively mitigates domain shift and leverages unlabeled data, leading to improved pedestrian identification performance.
• This approach demonstrates the potential of advanced generative models and self-training techniques for enhancing security and surveillance systems.