AeroDistinct: A generative data construction pipeline and benchmark for drone–bird discrimination in urban air mobility

Reliable UAV–bird discrimination is essential for vision-based perception in Urban Air Mobility (UAM), yet existing datasets provide limited co-annotated coexistence scenes and little controlled coverage of adverse weather. We present AeroDistinct, a benchmark and generative data-construction pipeline that combines public UAV and bird imagery, Qwen-Image-based scene-preserving editing for coexistence and weather augmentation, and field-collected images for validation and testing. We evaluate eight detectors spanning single-stage, two-stage, Transformer-based, and state-space architectures. Averaged across the eight detectors, adding generated coexistence data increases mAP 50 on the field-collected Test-Real set from 47.6% to 65.9%, while the complete Train-Weather configuration further raises it to 70.9%. On the separate weather-edited Test-Weather benchmark, weather augmentation improves mean mAP 50 from 59.6% to 77.8% and mean m A P 50 − 95 from 25.9% to 38.2%. D-FINE achieves 79.3% mAP 50 on Test-Real and 85.0% on Test-Weather. Performance gains also persist under Dense Fog, an unseen synthetic weather category excluded from training, suggesting improved robustness within this controlled evaluation protocol. These results show that controlled generative augmentation can improve UAV–bird discrimination without modifying detector architectures, while AeroDistinct provides a reproducible basis for evaluating data-centric robustness in UAM perception.