CommonSketch

Dataset Summary

CommonSketch is a semantically annotated sketch dataset introduced in the paper SEA: Evaluating Sketch Abstraction Efficiency via Element-level Commonsense Visual Question Answering. The dataset contains 23,100 human-drawn sketches across 300 object classes. Each sketch is paired with a fine-grained caption and element-level commonsense annotations for evaluating sketch abstraction and semantic recognizability.

Dataset Structure

CommonSketch/
  data/
    train-00000-of-00012.parquet
    ...
  metadata.csv
  captions.csv
  vqa_element_annotation.json
  metadata/
    classes.csv
    classes.txt
    commonsense_elements.json

Data Fields

The main row-level metadata is provided in metadata.csv.

Field	Description
file_name	Relative path to the sketch image.
caption	Fine-grained caption describing the sketch.
class_name	Object class name.
class_id	Class identifier from 001 to 300.
category	High-level category for the class.

The Parquet files under data/ provide the dataset examples for direct loading with the Hugging Face datasets library. Each row contains the following fields:

Field	Description
image	Sketch image embedded in the dataset.
file_name	Original relative image path.
caption	Fine-grained caption describing the sketch.
class_name	Object class name.
class_id	Class identifier from 001 to 300.
category	High-level category for the class.
element_annotation	JSON-serialized binary element annotation for the sketch.

The captions.csv file provides the image-caption pairs with the following fields.

Field	Description
image	Image file name.
caption	Fine-grained caption describing the sketch.

Annotation Files

vqa_element_annotation.json contains image-level binary element annotations. Each annotation indicates whether a class-specific commonsense element is present in the sketch.

metadata/commonsense_elements.json defines the class-level commonsense element schema used by the VQA annotations. Each class entry includes its class_id, total_elements, and the list of element definitions.

metadata/classes.csv provides the 300 class names, class IDs, high-level categories, and the number of commonsense elements per class. metadata/classes.txt provides the class list only.

Category Statistics

Category	# Classes
animal	61
body part	7
clothing	8
container	9
electronic device	22
food	28
furniture	25
icon	13
musical instrument	11
nature	14
sports equipment	14
structure	28
tool	37
vehicle	23

Usage

After upload to the Hugging Face Hub, the dataset can be loaded with:

from datasets import load_dataset

dataset = load_dataset("ziiio/CommonSketch")

The image field contains the sketch image, and element_annotation contains the image-level element annotation as a JSON string.

License

CommonSketch is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. If you use CommonSketch, please cite the SEA paper.

Citation

@article{park2026sea,
  title={SEA: Evaluating Sketch Abstraction Efficiency via Element-level Commonsense Visual Question Answering},
  author={Park, Jiho and Choi, Sieun and Seo, Jaeyoon and Sohn, Minho and Kim, Yeana and Kim, Jihie},
  journal={arXiv preprint arXiv:2603.28363},
  year={2026}
}

Links

• Paper: https://arxiv.org/abs/2603.28363
• PDF: https://arxiv.org/pdf/2603.28363
• Code: https://github.com/zihos/SEA
• Project page: https://zihos.github.io/SEA

Note

CommonSketch includes a subset of the SketchDUO dataset. For more information about SketchDUO, please refer to https://huggingface.co/datasets/ziiio/SketchDUO.