Region Aware Transformer with Intra-Class Compact for Unsupervised Aerial Person Re-identification

Zhizhi Lu, Hongxu Chen, Jian-Huang Lai

Abstract: The task of unsupervised person re-identification (ReID) involves matching images of the same individual across non-overlapping fields of view captured by cameras without the use of manual labels. In recent years, with the rise of aerial perspectives enabled by unmanned aerial vehicles (UAVs), this task has been extended to the aerial domain. However, compared to ground-based fixed cameras, UAV cameras offer greater flexibility, leading to more severe pose and viewpoint variations of pedestrians in aerial perspective. This results in significant intra-class variances during the clustering process within unsupervised methods. Existing unsupervised ReID methods rarely focus on the significant intra-class variations caused by the UAV perspective. To address these issues, we propose a Region Aware Transformer with Intra-Class Compact for Unsupervised aerial person ReID. Recognizing the invariance of local features under severe distortions, our Region Aware Transformer integrates both global and local information to achieve more stable feature representations. Furthermore, to mitigate the issue of substantial intra-class disparities among similar samples, we devise a Cluster Compact Loss. This loss function penalizes samples that stray too far from their respective cluster centers, encouraging more compact clustering. Our method not only outperforms state-of-the-art (SOTA) performance on UAV ReID datasets under unsupervised conditions, but it also demonstrates outstanding performance on ground-based ReID datasets.