NC4K

The NC4K camouflaged-object test set — 4121 images with binary masks, as one test split, which is how the benchmark is published and scored.

from datasets import load_dataset

ds = load_dataset("nobg/NC4K", split="test")   # 4121 rows
ds[0]["image"]   # PIL, original resolution
ds[0]["mask"]    # PIL, the binary mask

Why this mirror exists

NC4K is the largest camouflaged-object test set and is published test-only: every paper that reports it reports it over all 4121 images. The upstream mirror PassbyGrocer/NC4K had split it 2884 / 618 / 619 into train / validation / test — a partition invented by the mirror, not present in the benchmark. Scoring its test split yields a number over 619 images that no published result is comparable to, while looking like a valid NC4K score.

This mirror concatenates the three back into one test split and drops the invented boundary. Image and mask bytes are bit-identical to the source (verified by SHA-256 on all 4121 rows of both columns, in order — nothing is decoded or re-encoded); gt is renamed to mask for uniformity with the other nobg sets, and the per-object instance column is dropped.

The source also ships a fourth, single-row valid-00000-of-00001.parquet that its own dataset config does not reference. It is not a 4122th image: the row is a 352×352 resize — the standard COD training resolution — whereas all 4121 real rows are at original resolution across 1 878 distinct sizes, none of them 352×352. It is a preprocessing artifact and is excluded.

Licensing

No license is declared — not by NC4K's authors and not by the upstream mirror. It is left unset here rather than guessed; check with the original authors before any use beyond research. Same posture as nobg/COD10K.

Citation

@inproceedings{lv2021simultaneously,
  title={Simultaneously Localize, Segment and Rank the Camouflaged Objects},
  author={Lv, Yunqiu and Zhang, Jing and Dai, Yuchao and Li, Aixuan and Liu, Bowen and Barnes, Nick and Fan, Deng-Ping},
  booktitle={CVPR},
  year={2021}
}