Rendering vector graphics into pixel arrays can result in significant memory costs or loss of information, as shown in above Figure 1. We propose the first large-scale chart-based vector graphics dataset focusing on VG-based Detection and Chart Understanding.
| Task | Dataset | Type | Source | # Chart Type | # Nums |
|---|---|---|---|---|---|
| ChartQA | DVQA | RG | Synthetic | 1 | 300,000 |
| FigureQA | RG | Synthetic | 5 | 100,000 | |
| PlotQA | RG | Synthetic | 3 | 224,377 | |
| LEAF-QA | RG | Synthetic | 4 | 250,000 | |
| Chart-to-Table | ICPR 2020 | RG | Synthetic & Real | 15 | 40,322 |
| ICPR 2022 | RG | Real | 15 | 36,183 | |
| VG Detection (YOLaT used) | SESYD-Floorplans | VG | Synthetic | - | 1,000 |
| SESYD-Diagrams | VG | Synthetic | - | 1,000 | |
| VG Detection & Chart-to-Table | VG-DCU(Ours) | VG | Synthetic & Real | 16 | 15,197 |
The currently available public vector graphics datasets are limited to the two small datasets indicated in the Table and lack the complexity necessary for the advancement of vector image detection. In contrast, our proposed dataset comprises over 10,000 vector charts utilizing diverse primitives with rich attributes.
-
Vega-Lite: a synthetic subset generated with scripts and fictional data
-
Plotly:a real-world subset drawn by users.
Dataset Split: We collect 10,682 synthetic and 4,515 real charts in the VG-DCU dataset, by default using 80% as the training set and 20% as the test set. We divide the training and test set so that objects from the same category are included in both the training and testing set.
| Vega-Lite | Plotly | |||
|---|---|---|---|---|
| Chart Type | Train | Test | Train | Test |
| Area | 0 | 0 | 87 | 22 |
| Bar (Vert. & Hor.) | 2,400 | 601 | 1,704 | 426 |
| Box (Vert. & Hor.) | 0 | 0 | 482 | 121 |
| Donut&Pie | 2,395 | 599 | 200 | 50 |
| Line | 3,749 | 938 | 146 | 37 |
| Scatter | 0 | 0 | 640 | 161 |
| Heatmap | 0 | 0 | 154 | 39 |
| Counter | 0 | 0 | 180 | 45 |
| Violin | 0 | 0 | 82 | 21 |
| Sankey | 0 | 0 | 21 | 6 |
| Total | 8,544 | 2,138 | 3,609 | 906 |
Dataset Analysis
(a) The distribution map of categories and number of bbox instances. (b) The width-to-height ratio distribution of class and box instances
We will provide both Baidu Drive and Google Drive for downloading.
-
Vega-Lite: Google Drive
-
Plotly: Google Drive
BibTex:
@inproceedings{yolat24,
title={{Hierarchical Recognizing Vector Graphics and A New Chart-based Vector Graphics Dataset}},
author={Shuguang Dou, Xinyang Jiang, Lu Liu, Lu Ying, Caihua Shan, Yifei Shen, Xuanyi Dong, Yun Wang, Dongsheng Li, Cairong Zhao},
booktitle={IEEE Transactions on Pattern Analysis and Machine Intelligence},
volume={},
number={},
pages={},
year={2024}}
Please do consider 🌟 star our project to share with your community if you find this repository helpful!