Unsupervised Domain Adaptation for Person Re-Identification

Abstract

Person re-identification (ReID) is a fundamental computer vision task with numerous real-world applications, such as surveillance and robotics, enabling reliable human identification across several surveillance camera feeds in the fast-developing environment of computer vision. However, the performance of ReID models often degrades significantly when deployed in new camera domains due to the domain shift problem. To address this challenge, this thesis presents the Unified and Elevated Learning (UEL) framework for unsupervised cross-domain person ReID. that tackles the critical issue of domain shift. Domain shift occurs when a ReID model trained on one camera domain experiences a significant drop in performance when deployed in a new, visually distinct camera domain. This problem hinders the practical application of ReID systems in real-world scenarios. The UEL framework leverages a three-stage training approach that includes GAN-based camera-style data augmentation, source domain pre-training, and end-to-end cooperative learning. The key innovations of the UEL framework include adversarial training to extract camera-invariant features, cooperative learning to reduce noise in pseudo-labels, and dynamic fine-tuning of the data augmentation ratio. Extensive experiments on widely used benchmarks, such as Market-1501, DukeMTMC-reID, and MSMT17, demonstrate the superiority of the UEL framework over state-of-the-art unsupervised domain adaptation methods for person ReID. The ablation study further highlights the crucial role of the adversarial training component in the overall performance. The adaptability of the Unified and Elevated Learning framework to different backbone models is also validated. These results showcase the potential of the UEL framework for practical person ReID applications. The thesis contributes to the development of robust and adaptable person ReID systems that can operate effectively in real-world scenarios.