FineAir: Finest-grained Airplanes in High-resolution Satellite Images

Murat Osswald, Louis Niederloehner, Sascha Koejer, Tobias Ziedorn, Valerio Gulli, Michael Mommert, Helmut Mayer; Proceedings of the Winter Conference on Applications of Computer Vision (WACV) Workshops, 2025, pp. 1191-1199

Deep learning models require precisely categorized objects for the detection of small objects in high-resolution satellite images. Coarse-grained objects such as airplanes could be easily and accurately annotated. However determining fine-grained objects for example Airbus 320 is more challenging. In this respect transponders are substantial sources. While they transmit data about an airplane's current state such as location arrival and departure time they also provide not only fine-grained class but even Finest-Grained Class (FtGC) for example A320-200. In this work we match the collection time of the satellite images with transponder data points to generate accurately annotated airplanes for "Dataset of finest-grained airplanes in satellite images with 30-cm spatial resolution" (FineAir). Our data set addresses the limitations of other popular data sets and annotates airplanes with their FtGC in high-resolution satellite images for the first time. In FineAir around 1350 instances are annotated with FtGC and various levels of classes are provided for more than 5500 airplanes in total. Furthermore we have developed a unique way to split FineAir into non-overlapping sets. Our approach ensures that image data and class distributions are preserved along the sets for sparsely placed objects in high-resolution images. In addition our empirical experiments show that annotations are substantially more accurate than those of other public datasets for the joint classes. Furthermore the evaluations of state-of-the-art models for the complete set of 20 fine-grained classes in FineAir demonstrate that our two-step approach detection followed by classification predominantly outperforms the corresponding one-step approaches. We release FineAir in satellitepy for academic usages.