SUIM-ReMask / README.md
kuan2's picture
Release exact SUIM1525/110 split and SHA256 manifest
b299709 verified
|
Raw History Blame Contribute Delete
1.34 kB
metadata
task_categories:
  - image-segmentation
size_categories:
  - 1K<n<10K

SUIM — ReMask protocols

Exact SUIM copy used by the ReMask underwater experiments: 1,525 train images and 110 test images, with all corresponding masks and eight semantic classes. The original files are preserved. The 1,745 training-image/test-image/test-mask files are byte-identical to the frozen baseline input; per-file SHA256 and the archive hash are in manifest.json.

Download

Authenticate with an account that has access to this private repository (hf auth login), then run from the ReMask checkout:

hf download kuan2/SUIM-ReMask SUIM.zip --repo-type dataset --local-dir data
python -m zipfile -e data/SUIM.zip data
python baselines/run.py --method all --protocol all --data data/SUIM

Archive layout: SUIM/{train_val,TEST}/{images,masks}/. Train masks are provided for completeness; the SFUDA launcher excludes them from its adaptation data view. Labeled test masks are used for evaluation and checkpoint selection.

Attribution

Original dataset: Semantic Segmentation of Underwater Imagery: Dataset and Benchmark, Islam et al., IROS 2020. This repository is an experimental data copy, not a new dataset. Refer to the original authors for usage terms and citation.