Learning Invariance From Generated Variance for Unsupervised Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised representation learning is crucial for person re-identification (ReID).
• Current self-supervised contrastive learning methods rely on data augmentation, which can distort identity features.
• Existing Generative Adversarial Network (GAN)-based ReID methods often focus solely on id-unrelated features.

Purpose of the Study:

• To develop a novel data augmentation strategy for unsupervised person ReID using GANs.
• To improve representation learning by generating augmented views that better preserve identity-specific features.
• To achieve state-of-the-art performance in unsupervised person ReID.

Main Methods:

• Proposed a 3D mesh guided person image generator to disentangle identity-related and identity-unrelated features.
• Utilized GAN-based augmentation on both identity-unrelated (pose, camera style) and identity-related features.
• Introduced specific contrastive losses tailored for identity-related and identity-unrelated augmentations.
• Jointly trained generative and contrastive learning modules.

Main Results:

• Achieved new state-of-the-art performance in unsupervised person ReID.
• Demonstrated the effectiveness of GAN-based augmentation on both id-unrelated and id-related features.
• Validated the method on mainstream large-scale person ReID benchmarks.

Conclusions:

• The proposed GAN-based augmentation approach significantly enhances unsupervised representation learning for person ReID.
• Disentangling and augmenting both identity-related and identity-unrelated features leads to superior performance.
• This method offers a promising direction for improving ReID systems without labeled data.