Parallax_AI4QC

This dataset was used in the AI4QC project (Artificial Intelligence for Quality Control), in the context of parallax detection through an object detection task. It consists of a set of labeled parallax artefacts. This effect appears as a colorful pattern in images where cloud/heavy haze is present. Bounding boxes were defined around parallax artefacts in 1764 Sentinel-2 true colour images (jpg format). A parallax artefact appears mostly over clouds, although it can also be found over land. The bounding boxes are associated to one out of four labels:

0 : Parallax - Cloud (9529 bounding boxes)1 : Parallax - Land (245 bounding boxes)2 : Other Anomaly (4 bounding boxes)3 : Missing Data - Land (2 bounding boxes)

The label files are available in COCO format (json files). Two sets of labels are available: rotated_bboxes and straight_bboxes. This comes from the fact that parallax artefacts are often rotated on the image. The "rotated_bboxes" folder contains labels with an additional attribute which is a rotation value. These bounding boxes will contain only the rotated parallax artefact and not the background. However since most machine learning models don't account for a rotation value, straight bounding boxes were created as well. These are found in the "straight_bboxes" folder and the bounding boxes will contain some of the background as well as the parallax. 

One can combine the label files with the S2 images to train object detection algorithms to automatically detect parallax effects in a satellite image. A predefined train/test split is available (80% training and 20% testing), with training and testing zip folders containing the images and labels for each subset. The data was split according to 2 criterias: parallax over cloud vs land and geographic location.