Efficient Lightweight Multi-Source Domain Adaptation for Person Re-ID via Self-Paced Meta Learning

Person re-identification (Re-ID) aims to match individuals across different cameras, a task complicated by variations in camera positions, resolutions, and lighting conditions. While supervised training improves Re-ID model accuracy, it requires significant annotation efforts. Unsupervised domain adaptation (UDA) methods address this by leveraging unlabeled target domain data but often fail to fully utilize multiple source domains and are constrained by computational resources. This paper introduces a lightweight multi-source domain adaptation method for person Re-ID that combines meta-learning with pseudo-label-based UDA. By employing self-paced meta-learning and style enhancement techniques, the model learns domain-invariant knowledge from easy to difficult source domains, enhancing pseudo-label quality during adaptation. Our approach, based on an omni-scale feature extraction network using deep separable convolution, combines global and partial feature branches to capture richer pedestrian features. Experiments on public and real-world datasets demonstrate that our method achieves competitive performance with significantly fewer parameters and FLOPs compared to state-of-the-art models, proving its effectiveness and practicality.