Building Visual Anomaly Dataset from Satellite Data using ADS-B

In this paper, we present a novel way of obtaining extremely challenging image dataset for the purpose of benchmarking image anomaly detection methods. By definition, anomalies are rare occurrences, and therefore, annotation of anomalies using human workforce is difficult and costly, as large amounts of mostly non-anomalous data need to be checked. To alleviate this problem, we use satellite images from Planet.com as the source of visual data, and combine them with ADS-B data to detect airplanes in a semi-automatic way. This way, our definition of anomaly is an appearance of an airplane on mostly airplane-free images. This not only speeds up annotation, but also provides the exact specification of what constitutes an anomaly, in an objective way. The resulting meta-dataset, containing references to Planet.com imagery and accurate annotations will be published in the near future. It will include locations of nearly 100 positions of airplanes on satellite images and the corresponding references to satellite images, captured in vicinity of large airports in different parts of the world, in different climate zones.