--- license: other license_name: mixed-open task_categories: - image-segmentation configs: - config_name: default data_files: - split: train path: data/train-*.parquet --- # ReefSupport group sample 4854 reef images with semantic-segmentation masks, sampled from [reefsupport/marine-data](https://huggingface.co/datasets/reefsupport/marine-data) at revision `66f9614570767668fdaa654c5881c8bfc20f9a16` with seed 42 by `build_sample_dataset.py`. | source | images | mask labels | licence | |---|---|---|---| | `coralscapes-v2` | 2427 | coralscapes-v2-95 (1-95 = upstream class ids, 0 = unlabelled) | Apache-2.0 | | `reef-support-seaview-labels` | 2427 | seaview-hc-sc (1 = hard coral, 2 = soft coral, 0/255 = ignore) | images CC-BY-3.0, masks CC-BY-4.0 | Everything is in one `train` split; the upstream split of each image is kept in `source_split`. The two sources use different mask label sets (see `mask_label_set`), so harmonise them before training a single model. `manifest.csv` lists every image by `image_sha256`. Attribution is in `NOTICE`. ## Columns `image_sha256`, `source`, `source_split`, `image` (encoded image), `mask` (encoded single-channel PNG of class ids), `mask_label_set`, `licence`.