SAVAir: A Physics-Guided Temporal Synthetic Dataset for Satellite-Video Aircraft Detection and Tracking

The dataset contains 1,281 videos and 99,960 frames, combining real remote-sensing backgrounds with real and physics-guided synthetic aircraft instances. It provides horizontal bounding-box annotations and persistent track identities, together with oriented bounding boxes and visible-instance masks for synthetic aircraft, aircraft configuration and structural attributes, flight-phase information, sequence-level metadata, and auxiliary Chinese and English semantic descriptions.

The synthetic aircraft are generated using a physics-guided temporal synthesis pipeline incorporating structural modeling, illumination and material reflectance simulation, atmospheric radiative transfer, optical degradation, product-domain calibration, and trajectory constraints. SAVAir also provides benchmark annotations and protocols for aircraft detection and multi-object tracking.