TactileEval Dataset

TactileEval decomposes tactile-graphic quality into five BANA-aligned quality dimensions (view, parts, background, texture, line quality) across six object families, yielding 30 task families and 14,095 option-level annotations. Each record corresponds to a natural photo / tactile drawing pair and a specific quality option with majority-vote label, vote counts, and provenance metadata.

Repository layout

images/
  <Family>/<Object>/<Natural|Tactile>/*.{jpg,png}
processed/
  records_full.jsonl
  splits/{train,val,test}.jsonl
  family_splits/F{1..6}/{train,val,test}.jsonl
  dataset_summary.{csv,json}

JSONL fields:

Field	Description
pair_id	NaturalRel::TactileRel identifier
task_family	Task code (e.g., F1QL)
option_id	Option identifier (e.g., too_thick)
option_description	Plain-language description
natural_image, tactile_image	Paths relative to images/
votes_total, positives, negatives	Vote stats
label	Majority label (0/1)
vote_fraction	positives / total
status_counts	counts of approved/submitted ballots

Splits

• splits/train.jsonl (11,348 records)
• splits/val.jsonl (1,341 records)
• splits/test.jsonl (1,406 records)

family_splits/ mirrors these splits per object family.

Usage

from datasets import load_dataset

ds = load_dataset("Adnank1998/TactileEval", name="full", split="train")
family = load_dataset("Adnank1998/TactileEval", name="family_f1", split="train")
example = ds[0]
print(example["natural_image"], example["label"])

Images are stored under images/; join the relative path returned in natural_image/tactile_image with the local dataset root to load the files.

Available configurations:

• full: All families (default).
• family_f1 through family_f6: Per-family subsets matching the paper splits.

Each split lives in processed/, enabling the Hugging Face dataset viewer via the bundled dataset_infos.json.

Available configurations:

• full: All families (default).
• family_f1 … family_f6: Per-family subsets matching the paper splits.

Each split lives in processed/, so the Hugging Face dataset viewer can load the files directly via the bundled dataset_infos.json.

Citation

@misc{khan2026tactileevalstepautomatedfinegrained,
  title={TactileEval: A Step Towards Automated Fine-Grained Evaluation and Editing of Tactile Graphics},
  author={Adnan Khan and Abbas Akkasi and Majid Komeili},
  year={2026},
  eprint={2604.19829},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2604.19829}
}

Contact

Questions? Open an issue or email adnankhan5@cmail.carleton.ca.

Acknowledgements

This work was supported in part by MITACS and the Digital Alliance of Canada. We thank the student volunteers at the Intelligent Machines Lab (iML), Carleton University, for their contributions, and Joshua Olojede and Hoda Vafaeesefat for their help with the AMT annotation environment.