Coralscapes and CoralscapesV2

Coralscapes is a dense semantic segmentation dataset of coral reef imagery modelled on the Cityscapes urban dataset, built by Jonathan Sauder, Viktor Domazetoski, Guilhem Banc-Prandi, Gabriela Perna, Anders Meibom and Devis Tuia with EPFL's ECEO lab and the Transnational Red Sea Center. The first release has 2075 images, 39 classes and about 174,000 annotations from five countries, and was presented at the ICCV 2025 Joint Workshop on Marine Vision. CoralscapesV2 extends it to 2433 frames, 95 fine-grained classes, about 270,000 polygons and 65,738 fish instance masks from 55 Red Sea reef sites, with co-authors from institutions in Djibouti, Eritrea, Jordan and Sudan; it is a paper for the ECCV 2026 Workshop on Marine Vision (arXiv 2609.12826).

The datasets exist to train models that read underwater footage. They are the training data for the segmentation models in DeepReefMap, EPFL's Apache-licensed software that turns handheld video into a labelled 3D point cloud, an ortho-mosaic and benthic cover statistics for a reef.

Both load from the Hugging Face Hub with the datasets library. V2 is released under Apache 2.0; the V1 page does not state a dataset licence. V2's video sequences total around 206 GB, so streaming is advised. Pretrained segmentation models in 39- and 95-class versions are published alongside.