Learning to Adapt Invariance in Memory for Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation is crucial for person re-identification (re-ID) to bridge the gap between different datasets.
• Current methods often focus on reducing inter-domain shifts but neglect intra-domain relationships within the target domain.

Purpose of the Study:

• To propose a novel unsupervised domain adaptation framework for person re-ID.
• To investigate and leverage three types of invariance: Exemplar-Invariance, Camera-Invariance, and Neighborhood-Invariance.
• To improve the performance of person re-ID across different domains.

Main Methods:

• Introduced an exemplar memory to store sample features for enforcing invariance constraints globally.
• Developed a Graph-based Positive Prediction (GPP) method to identify reliable neighbors in the target domain.
• The framework integrates Exemplar-Invariance, Camera-Invariance, and Neighborhood-Invariance for robust adaptation.

Main Results:

• Demonstrated that the three invariance properties are complementary and essential for effective domain adaptation.
• Highlighted the critical role of the exemplar memory in learning invariances with minimal computational overhead.
• Showcased that GPP significantly enhances invariance learning and overall performance.

Conclusions:

• The proposed framework achieves state-of-the-art adaptation accuracy on large-scale re-ID benchmarks.
• The integration of exemplar memory and GPP offers an effective approach to unsupervised domain adaptation in person re-ID.
• The findings underscore the importance of considering intra-domain variations for successful knowledge transfer.