ziiio/CommonSketch

---
pretty_name: CommonSketch
license: cc-by-4.0
task_categories:
- image-classification
- image-to-text
- visual-question-answering
language:
- en
tags:
- image
- sketch
- captions
- visual-question-answering
- commonsense
- abstraction
- computer-vision
size_categories:
- 10K<n<100K
---

# 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](https://arxiv.org/abs/2603.28363). 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

```text
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:

```python
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

```bibtex
@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.

数据集名称：CommonSketch
许可证：CC BY 4.0
任务类别：
- 图像分类（image-classification）
- 图像转文本（image-to-text）
- 视觉问答（visual-question-answering）
语言：
- 英语
标签：
- 图像（image）
- 草图（sketch）
- 标题（captions）
- 视觉问答（visual-question-answering）
- 常识（commonsense）
- 抽象性（abstraction）
- 计算机视觉（computer-vision）
样本规模：10K<n<100K

# 通用草图数据集（CommonSketch）

## 数据集概览
通用草图数据集（CommonSketch）是一篇发表于论文《SEA: Evaluating Sketch Abstraction Efficiency via Element-level Commonsense Visual Question Answering》（https://arxiv.org/abs/2603.28363）的语义标注草图数据集。本数据集包含覆盖300个目标类别的23100幅人工绘制草图。每幅草图均配有细粒度标题与元素级常识标注，用于评估草图的抽象性与语义可识别性。

## 数据集结构
text
CommonSketch/
data/
train-00000-of-00012.parquet
...
metadata.csv
captions.csv
vqa_element_annotation.json
metadata/
classes.csv
classes.txt
commonsense_elements.json

## 数据字段
行级元数据主要存储于`metadata.csv`中，字段说明如下：

| 字段 | 说明 |
| --- | --- |
| `file_name` | 指向草图图像的相对路径 |
| `caption` | 描述草图的细粒度标题 |
| `class_name` | 目标类别名称 |
| `class_id` | 类别标识符，取值范围为001至300 |
| `category` | 该类别的高级类别归属 |

`data/`目录下的Parquet文件可通过Hugging Face `datasets`库直接加载，其中每行包含以下字段：

| 字段 | 说明 |
| --- | --- |
| `image` | 嵌入数据集的草图图像 |
| `file_name` | 原始图像相对路径 |
| `caption` | 描述草图的细粒度标题 |
| `class_name` | 目标类别名称 |
| `class_id` | 类别标识符，取值范围为001至300 |
| `category` | 该类别的高级类别归属 |
| `element_annotation` | 用于描述草图的JSON序列化二值化元素标注 |

`captions.csv`文件存储图像-标题配对数据，字段说明如下：

| 字段 | 说明 |
| --- | --- |
| `image` | 图像文件名 |
| `caption` | 描述草图的细粒度标题 |

## 标注文件
`vqa_element_annotation.json`包含图像级二值化元素标注，每条标注用于指示草图中是否存在对应类别的常识元素。
`metadata/commonsense_elements.json`定义了VQA标注所使用的类别级常识元素模式，每个类别条目包含其`class_id`、总元素数（`total_elements`）以及元素定义列表。
`metadata/classes.csv`提供300个类别的名称、类别ID、高级类别归属以及每个类别的常识元素数量；`metadata/classes.txt`仅存储类别列表。

## 类别统计
| 类别 | 类别数量 |
| --- | ---: |
| 动物 | 61 |
| 身体部位 | 7 |
| 服饰 | 8 |
| 容器 | 9 |
| 电子设备 | 22 |
| 食品 | 28 |
| 家具 | 25 |
| 图标 | 13 |
| 乐器 | 11 |
| 自然事物 | 14 |
| 体育器材 | 14 |
| 建筑结构 | 28 |
| 工具 | 37 |
| 交通工具 | 23 |

## 使用方法
本数据集上传至Hugging Face Hub后，可通过以下代码加载：
python
from datasets import load_dataset

dataset = load_dataset("ziiio/CommonSketch")

其中`image`字段存储草图图像，`element_annotation`字段存储以JSON字符串形式呈现的图像级元素标注。

## 许可证
通用草图数据集（CommonSketch）采用知识共享署名4.0国际版（CC BY 4.0）协议发布。若使用本数据集，请引用SEA相关论文。

## 引用格式
bibtex
@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}
}

## 相关链接
- 论文：https://arxiv.org/abs/2603.28363
- PDF全文：https://arxiv.org/pdf/2603.28363
- 代码：https://github.com/zihos/SEA
- 项目主页：https://zihos.github.io/SEA

## 说明
通用草图数据集（CommonSketch）包含SketchDUO数据集的子集。如需了解SketchDUO的更多信息，请访问https://huggingface.co/datasets/ziiio/SketchDUO。