Deucalion: A dataset for flood related research.

Current paper, introduces Deucalion, a dynamic dataset currently consisted of 5,042 photos from Instagram, which was scraped for researching on the medicane Ianos during September 2020, and two Kaggle sources with flood related content. Other sources, including internet search engines and Flickr complete the data provider list of current version. 1,664 photos of them were imported in LabelStudio and objects were identified and digitized. The objects extracted are currently classified in 15 different classes, including flood, sea, pools, rocks and mud, vegetation. The entire dataset was used for fine-tuning a vgg19, providing thus SOTA metrics, while various classes of the 1,664 subset were used to train a YOLO model. The plethora of different classes and sources, the real world captures, along with the dynamic nature of Deucalion is expected to emerge it as a significant research dataset. The dataset currently can be available upon request.