COSMOS-ReID: Benchmarking Open-World Cross-Sensor Object Re-Identification Across Satellite and UAV Imagery

Cross-modal object re-identification (ReID) is challenging because sensor-dependent appearances differ substantially. We introduce COSMOS-ReID, an open-world cross-sensor benchmark for satellite and UAV imagery spanning RGB-SAR and multispectral-SAR vessels and RGB-LWIR vehicles. Our contributions are: (a) a unified open-world evaluation protocol spanning strict directional cross-modal retrieval, mixed-modality galleries, unknown-identity rejection, and online identity discovery; (b) ShipCross-ID, a new AIS-linked multispectral-SAR vessel dataset connecting FUSAR SAR and Sentinel-2 observations; and (c) controlled experiments examining how cross-modal pretraining, sensing-modality gap, and training-set size affect retrieval and open-set performance. Across settings, DINOv3 performs best most often. Existing CLIP- and PMSL-style objectives improve retrieval without improving open-set decisions, showing that strong retrieval alone is insufficient for reliable open-world ReID. COSMOS stands for Cross-sensor Open-Set Multimodal Object Re-identification for UAV and Satellite imagery.