UAVDB: Point-Guided Masks for UAV Detection and Segmentation

This version of UAVDB extends the original dataset by adding high-quality instance segmentation masks generated using SAM2, guided by PIC-based bounding box prompts. While retaining the original 18,061 RGB images across training (10,763), validation (2,720), and test (4,578) splits, it enhances spatial annotation granularity to support dense supervision. This upgrade enables more accurate training and evaluation of UAV perception algorithms across diverse viewpoints and complex aerial scenes.