aviadcohz/Detecture_ICLR_Benchmarking

Detecture ICLR Benchmarking

The four evaluation routes reported in the ICLR 2027 submission on sub-semantic image segmentation, bundled so the published numbers can be reproduced from a single download.

This is a smaller, paper-aligned release. An earlier bundle, aviadcohz/Detecture_Benchmarking, carried five datasets at 4.3 GB for a previous version of this work. That one included routes the current paper does not report. This release carries only what the paper evaluates on.

Directory	Paper label	Images	Regions / image	Source
RWTD/	RWTD	253	2	Real-world texture photographs
CAP/	RWTD-COCO	256	2	COCO-Stuff, by deterministic annotation reuse
ADE20k_Detecture/	TextureADE	212	3.91 avg	Mined from the ADE20K validation split
CSTD/	CSTD	256	2 by construction	Synthetic, SD-1.5 + ControlNet over DTD textures

The directory names are the on-disk names the evaluation code expects, not the paper's display labels. They are deliberately left unrenamed: clone this repo into ~/datasets/ and the evaluation configs resolve every route without edits. The table above is the mapping.

Layout

RWTD/  CAP/  ADE20k_Detecture/
├── images/              RGB images
├── masks/               per-image region maps
├── textures_mask/       per-texture binary masks, <id>_mask_<k>.png
├── overlays/            image + mask visualisations (not ground truth)
├── metadata.json        image paths, mask paths, descriptions
└── summary.json         per-dataset statistics

CSTD/
├── images/              256 generated images
├── textures_mask/       512 binary masks, two per image
├── metadata.json        the 256-image evaluation subset
├── verified_256_ids.json  the accepted ids
├── screen_scores.json   per-image screening metrics for all 10,000
├── screen_rank.json     the ranking those scores induce
└── screen_cstd.py       the screening script, so the selection is reproducible

Paths inside every metadata.json are relative to the repository root, so the bundle can be placed anywhere.

Read this before using CSTD

CSTD is redistributed here as a 256-image hand-verified subset, not as published. The original release is on Kaggle as architexanonymous/cstd-controlnet-synthetic-texture.

The reason is a ground-truth problem. CSTD's released regions/*.png is the stitching mask fed into ControlNet, not an annotation of what came out. Where the generator invented a third material or drifted from the mask, the ground truth silently stops describing the image. Three failure modes were observed and confirmed by eye: a region containing two distinct textures, a third material appearing at an edge or corner, and a contour that does not sit on any real appearance change.

All 10,000 images were therefore screened on texture features, ranked, and the top candidates reviewed by eye. 274 of 1,296 reviewed candidates were accepted, a 21% pass rate, and the top 256 form this subset. screen_cstd.py and screen_scores.json are included so the screening is reproducible rather than asserted.

Anyone evaluating on CSTD as published will get different numbers, and should.

Download

cd ~/datasets
git lfs install
git clone https://huggingface.co/datasets/aviadcohz/Detecture_ICLR_Benchmarking .

Or from Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="aviadcohz/Detecture_ICLR_Benchmarking",
    repo_type="dataset",
    local_dir="~/datasets",
)

Provenance

RWTD-COCO carries no predicted pixels. A whitelist of 28 surface-like COCO-Stuff classes proposes adjacent label pairs, crops are enumerated around the shared boundary and scored by a closed-form structural criterion, and every ground-truth pixel is a deterministic remap of existing human annotation. No SAM, CLIP, DINO or saliency model participates in its construction.

TextureADE is mined from the natural ADE20K validation split by a geometry-first scoring procedure: acceptance rests on observable mask geometry, and a frozen vision-language annotator is queried only afterwards, against a region that has already been accepted.

RWTD is redistributed unchanged as a cross-domain stress test.

Evaluation protocol

Every number in the paper comes from one protocol, applied identically to every method and route: 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. Region counts are inferred, never supplied.

Licence

CC-BY-4.0 for this bundle. Upstream corpora keep their own licences: ADE20K, COCO-Stuff and DTD are each governed by their original terms, and this release does not relicense them.