Planomaly (short for PLAnes as an aNOMALY) is a rare anomalies dataset, intended for study of anomaly detection algorithms and the corresponding evaluation methods. The final dataset will be comprised of large satellite images (approximately 8000 x 4000 pixels in size), which contain commercial airplanes in flight (approximately 30 x 30 pixels in size).
The specific task, posed by this dataset is: “Find airplanes by training only on non-airplane data”. This task is exceedingly difficult, and requires the use of a generative model.
The dataset is published in a form of metadata. We provide unique image IDs from Planet.com, but the individual researchers should arrange with Planet.com for access to images. The best way to do it is via Planet.com’s Education and Research Program. In addition, we provide exact location of airplane(s) in each image in pixel coordinates.
Finally, please note that versioning of this dataset does matter. We will release PlanomalyV0 shortly. This will be a single image, containing three airplanes. Due to large size of the image, this yields 12,000+ training image patches, 1,200 normal testing patches, and 3 anomalous patches containing airplanes.
Full dataset, with about 100 anomalous patches and almost as many images will be released in early 2020. Please contact Janez Perš for any additional info.