Unsupervised group re-identification from aerial perspective via strategic member harmonization

Abstract

Group re-identification (G-ReID) aims to match group images of the same identity. Existing G-ReID methods perform well on ground-based datasets, but remain unexplored in aerial perspective. One reason is the significant human effort required for aerial associations and the inability of unsupervised methods to address low-quality aerial pedestrian detection and reduced feature visibility. To address these issues, we propose Strategic Member Harmonization. Strategic members are harmonized to complement potential information lost or destroyed due to low-quality detections or significant member variations, thus forming harmonization groups. Harmonization groups introduce a richer layer of the underlying information, mitigating clustering inaccuracies gradually. To address the lack of aerial G-ReID datasets, we construct a new aerial dataset with 10,168 group images and 653 different group identities. Our approach achieves state-of-the-art performance on our dataset and performs well on other ground-based datasets. Our dataset is available at https://github.com/chen1hx/UAV-Group.