Abstract: The rapidly increasing use of unmanned aerial vehicles (UAVs) for surveillance has paved the way for advanced image analysis techniques to enhance public safety. Among many others, person re-identification (ReID) is a key task. However, much of the current literature is centered on research datasets, often overlooking the practical challenges and unique requirements of UAV-based aerial datasets. We close this gap by analyzing these challenges, such as viewpoint variations and lack of annotations, and proposing a framework for aerial person re-identification under unsupervised setting. Our framework integrates three stages: generative, contrastive, and clustering, designed to extract view-invariant features for ReID without the need for labels. Finally, we provide a detailed quantitative and qualitative analysis on two UAV-based ReID datasets, and demonstrate that our proposed model outperforms state-of-the-art methods with an improvement of up to 2% in rank-1 scores.
Loading