RWTD-COCO

Single natural appearance transitions built from COCO-Stuff by deterministic reuse of human annotation.

One of the four evaluation routes in the ICLR 2027 submission on sub-semantic image segmentation: partitioning an image into regions that are coherent in appearance and describable in language, but that need not correspond to any object, part or material class.

• Images: 256
• Code: github.com/aviadcohz/Qwen2SAM_Detecture_Benchmark
• Weights: aviadcohz/Detecture-ICLR-2027
• All four routes in one download: aviadcohz/Detecture_ICLR_Benchmarking

Layout

CAP/
├── images/            RGB images
├── textures_mask/     per-texture binary masks, <id>_mask_<k>.png
├── metadata.json      image paths, mask paths, descriptions
└── summary.json       dataset statistics

The three real-world routes also carry masks/ and overlays/; overlays are visualisations, not ground truth.

The directory inside this repo is named CAP rather than RWTD-COCO, because that is the name the evaluation configs resolve (fairness_baseline_suite/src/paths.py). Paths inside metadata.json are relative to the repository root, so the folder can be placed anywhere.

Use

cd ~/datasets
git lfs install
git clone https://huggingface.co/datasets/aviadcohz/RWTD-COCO
mv RWTD-COCO/CAP . && rm -rf RWTD-COCO

Then, from the benchmark repo:

cd Qwen2SAM_Detecture_Benchmark/fairness_baseline_suite
PYTHONPATH=src python src/run_fairness.py --model detecture --dataset RWTD_COCO

Evaluation protocol

Every number reported on this route comes from one protocol applied identically to every method: no ground-truth region count in the prompt, no inverse-mask completion, no truncation of proposals to a known count, and no dropping of images where a method returns nothing. The region count is inferred, never supplied. Results obtained this way are not comparable to evaluations that supply it.

Provenance

No predicted pixel enters the ground truth. A whitelist of 28 surface-like COCO-Stuff classes proposes adjacent label pairs; a pair becomes a candidate only where each class covers at least 2.5% of the map and the two share at least 12 four-neighbour edges, so the candidate exists precisely because annotators placed them on opposite sides of a boundary. Crops are enumerated around that shared boundary and scored by a closed-form structural criterion. No SAM, CLIP, DINO or saliency model participates; RGB pixels are read only after selection.

Licence

CC-BY-4.0 for this packaging. Upstream corpora keep their own terms: COCO-Stuff.