Qwen-camo

CAMO images and masks, plus a short concept name per image generated by Qwen/Qwen3.6-27B-FP8, so prompt-conditioned segmentation on camouflaged subjects needs one load_dataset call instead of a generic placeholder prompt. 1250 rows, 2 splits.

from datasets import load_dataset

ds = load_dataset("nobg/Qwen-camo", split="test")
ds[0]["image"]          # PIL, original resolution, byte-identical to nobg/camo
ds[0]["mask"]           # PIL, the binary mask
ds[0]["prompt"]         # 'a leaf-tailed gecko'
ds[0]["prompt_tokens"]  # 7 -- CLIP BPE count, guaranteed <= 32

Why this exists

Camouflage is the hardest case for an open-vocabulary segmenter, because the subject is by construction hard to name — and it was the case with no name available. In nobg's SAM 3 training mix, CAMO carried an empty prompt field, meaning all 1250 images fell back to one run-level phrase ("the main foreground subject"). This dataset supplies the per-image concept instead.

How the prompts were made

Each image was composited subject-only onto a flat neutral grey background and shown to Qwen/Qwen3.6-27B-FP8 with an instruction asking for a 1–4 word noun phrase. Greedy decoding (temperature=0), so the annotation is reproducible.

The compositing detail matters and is easy to get wrong: a straight-alpha RGBA cutout is a no-op for a vision-language model, because every image processor calls .convert("RGB") internally, which discards the alpha and restores the original pixels. The model would then describe the unmasked scene — for camouflage, that means naming the habitat instead of the animal, plausibly enough that nothing looks wrong. The background is therefore explicitly pasted, and the build verified this by annotating a sample both ways and confirming the two disagree.

column	type	contents
prompt	string	the cleaned concept name, ≤ 32 CLIP tokens. "" where the model refused or named nothing usable
prompt_tokens	int32	measured CLIP BPE count, so the bound is auditable from inside the data
prompt_raw	string	the model's verbatim output, before normalization

Token counts: median 5, max 11, budget 32. 592 distinct concepts over 1250 prompts. 0 rows have an empty prompt.

An empty prompt is a deliberate fallback value, not a hole: nobg's loader treats it as "use the run-level prompt", so a refusal degrades to the previous behaviour instead of training on a wrong name.

Limitations — read before training on this

• The prompts are machine-generated and unreviewed. No human checked 1250 annotations. Expect errors, especially in species-level naming.
• They describe the masked cutout, not the scene. For camouflage this can differ from what a human naming the photograph would say: a subject that is only recognizable in context may be named more generically once isolated.
• Images and masks are byte-identical to nobg/camo — no re-encode or resize. Only the two 400×300 overlaid_mask_* thumbnail columns were dropped.

Licensing

cc-by-nc-4.0, inherited from nobg/camo — non-commercial. Qwen/Qwen3.6-27B-FP8 is Apache-2.0, so the generated text carries no additional restriction, but the images govern.

These prompts are unrelated to the FlowDIS paper; the name follows the <method>-<source> convention of nobg/FlowDIS-DIS5K.

Citation

Images and masks — CAMO:

@article{le2019anabranch,
  title={Anabranch network for camouflaged object segmentation},
  author={Le, Trung-Nghia and Nguyen, Tam V and Nie, Zhongliang and Tran, Minh-Triet and Sugimoto, Akihiro},
  journal={Computer Vision and Image Understanding},
  volume={184},
  pages={45--56},
  year={2019}
}