Multi-Domain Adversarial Feature Generalization for Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Fully-supervised person re-identification (Person Re-ID) models excel on single datasets but degrade on new camera networks.
• Existing cross-dataset domain adaptation methods require unlabeled target domain data, hindering immediate deployment.
• Practical Person Re-ID systems need to function immediately on new sites without retraining.

Purpose of the Study:

• To develop a Person Re-ID system that generalizes to unseen camera systems without requiring target domain data.
• To reformulate Person Re-ID as a multi-dataset domain generalization problem for enhanced scalability and practicality.

Main Methods:

• Proposed a multi-dataset feature generalization network (MMFA-AAE) for learning universal, domain-invariant feature representations.
• Utilized an adversarial auto-encoder architecture.
• Incorporated the Maximum Mean Discrepancy (MMD) measure to align feature distributions across multiple domains.

Main Results:

• The MMFA-AAE approach demonstrated significant effectiveness in generalizing Person Re-ID to unseen camera systems.
• Outperformed existing domain generalization Person Re-ID methods.
• Surpassed state-of-the-art supervised and unsupervised domain adaptation methods.

Conclusions:

• The proposed MMFA-AAE enables immediate deployment of Person Re-ID systems on new sites without data collection or retraining.
• Achieved a universal domain-invariant feature representation for robust Person Re-ID performance across diverse camera networks.
• Offers a practical and scalable solution for real-world Person Re-ID applications.