Complex Optical Remote-Sensing Aircraft Detection Dataset and Benchmark

Aircraft detection in remote-sensing images is significant in both military and civilian fields, such as air traffic control and battlefield dynamic monitoring. Deep-learning (DL) methods can achieve promising detection performance with sufficient and labeled samples. However, current aircraft datasets are mainly from a single data source and lack diverse scenes and targets, making it difficult to train a robust and generalized detector. Therefore, we manually label and construct a complex optical remote-sensing aircraft target detection dataset from Google Earth and multiple satellites such as WorldView-2, WorldView-3, Pleiades, Jilin-1, and IKONOS. It contains 7337 images covering typical airports and various rare scenes, including the aircraft carrier, ocean, and land with flying aircraft. The dataset consists of 32285 civil and military aircraft instances, including bombers, fighters, and early warning aircraft. These targets range from
$4\times $
4 to
$240\times240$
pixels and are all labeled with both horizontal bounding box (HBB) and oriented bounding box (OBB) annotations. The various scenes and sufficient instances can fully support the training and evaluation of data-driven algorithms. Meanwhile, based on the constructed dataset, we train and evaluate several detectors to provide a benchmark and help promote the development of aircraft detection techniques.