CoralscapesV2: Panoptic Segmentation of Red Sea Coral Reefs with 95 Benthic Classes and Fish Instances
- Provider
- Sauder J., Tuia D. et al. (EPFL ECEO)
- License
- Apache-2.0 (source)
- Size
- 524 files
- Format
- Parquet (in sampled files)
- Stored at
- Hugging Face (https://huggingface.co/datasets/josauder/coralscapesV2)
- Access
- Free and public. No account with the data provider is needed.
- Category
- Environmental Science
- Source
- https://arxiv.org/abs/2609.12826
About this dataset
CoralscapesV2 is a scene-understanding dataset for coral reefs and the first for panoptic segmentation underwater (ECCV 2026 Workshop on Marine Vision, arXiv 2609.12826), extending Coralscapes (EPFL ECEO). It holds 2,433 densely annotated video frames at 1024x2048 px from 41 dive sites in the Red Sea region, split by site into 1,790 train, 170 validation and 473 test images. Semantic masks use 95 fine-grained benthic and visual categories (live, bleached and dead hard and soft coral genera, seagrass, algae, sand, rock, trash, fish and others), each mapped onto the 39 original Coralscapes classes so older models still evaluate. An instances configuration adds 65,738 exhaustive fish instance masks annotated with video context, and a sequences configuration returns 31-frame clips around each labelled anchor frame for temporal models. Measured: 524 Parquet and metadata files, about 220 GB. Load with datasets.load_dataset("josauder/coralscapesV2"). Licensed Apache 2.0; cite CoralscapesV2 and Coralscapes (Sauder et al.).
How to load this dataset in Python
from datasets import load_dataset
ds = load_dataset("josauder/coralscapesV2")
print(ds)Data structure
File types in sampled files: Parquet
Files
Example file: data/test-00000-of-00003.parquet
Analyze this dataset without downloading it
On Scigantic, CoralscapesV2: Panoptic Segmentation of Red Sea Coral Reefs with 95 Benthic Classes and Fish Instances opens in a hosted JupyterLab notebook with the data already mounted, so you can read files where they are stored instead of copying them first. Each notebook includes an AI assistant that can see the mounted files and write and run analysis code. Open the dataset page to launch a notebook, or browse the full catalog.
This page is also available as Markdown and as JSON.
Related Environmental Science datasets
- Alaska Small-Mesh (Shrimp) Trawl Survey Catch Table, 1953 onward (NOAA AFSC and ADF&G)
- CanadaFireSat: High-Resolution Multimodal Wildfire Forecasting Benchmark for Canada
- Global 100m Projections of Biodiversity Intactness for the years 2017-2025 (v1.1)
- WRI/Google DeepMind Global Drivers of Forest Loss (1km)
- Steller Sea Lion Diet from Scat Hard Parts, Alaska Rookeries and Haulouts, 1990-2009
- GOBAI-O2 v2.3 Global Ocean Interior Dissolved Oxygen, Monthly 2004-2024 (Icechunk)
- WorldTensor GeoZarr: 757 harmonized Earth system variables on a 0.25 degree global grid, 1900-2025 (Rodriguez-Pardo and Tavoni, Taylor Geospatial, Source Cooperative)
- Nkhotakota Camera Traps (Malawi)