Pengyu965/ChartDete-PMC2022

Refined PMC chart element detection dataset (CACHED, ICDAR 2023)

This is the official dataset release of the ICDAR 2023 paper Context-Aware Chart Element Detection (CACHED). The paper was published at ICDAR 2023 (arXiv: May 2023). The data here is the original dataset from the 2023 paper, unchanged. It was uploaded to Hugging Face in October 2026 only to make it easier to download. Before that, it was available only as a Google Drive archive linked from GitHub. This is not a new dataset, a new version, or new work.

• 📄 Paper: Context-Aware Chart Element Detection (ICDAR 2023) · arXiv:2305.04151
• 💻 Code: github.com/pengyu965/ChartDete
• 🤖 Model: Pengyu965/ChartDete
• 📚 Dataset: Pengyu965/ChartDete-PMC2022 (this repo)

Pengyu Yan*, Saleem Ahmed*, David Doermann (*equal contribution). University at Buffalo.

Overview

This dataset contains real-world chart images from PubMed Central (PMC), annotated for chart element detection. It is built from the PMC chart dataset of the ICPR 2022 CHART-Infographics competition (Davila et al.). In the CACHED paper, we refined its element categories and converted the annotations to the COCO object detection format:

• Axis-related elements are split by axis. The original PMC annotations label axis titles, tick labels, and tick marks jointly. Here they are separated into x_title / y_title, xlabel / ylabel (tick labels), and x_tick / y_tick (tick marks).
• Structural areas are added. Besides the original plot_area, three area objects were generated with rule-based methods: x_axis_area, y_axis_area, and legend_area.
• 18 classes in total, covering the basic chart elements outside the data marks. The bar_detection subset adds a 19th class, bar.

The refinement and conversion code is in pmc2coco_converter/ of the GitHub repo.

What is in this repository

The same data is provided in three forms. Pick the one that fits your use:

Path	Size	Content	Use it for
pmc_coco.zip	304 MB	The pmc_coco/ folder: COCO JSON annotations and images, exactly as used by the CACHED configs	Training and evaluation with the ChartDete code (MMDetection)
pmc_2022.zip	2.1 GB	The full archive, identical to the original Google Drive release: pmc_coco/ plus the original ICPR 2022 PMC training and test releases (train_original/, test_original/)	A complete mirror, including the original competition annotations
parquet/	385 MB	The same images and annotations as pmc_coco/, as Parquet (two configs: element_detection, bar_detection)	The dataset viewer and datasets.load_dataset

The COCO data is stored as archives because it has about 8,000 files, and downloading them one by one would run into the Hub's request rate limits. pmc_coco.zip has exactly the same files as the pmc_coco/ folder in pmc_2022.zip. Unzipped, the full archive looks like this:

pmc_2022/
├── pmc_coco/                           # refined categories, COCO format (= pmc_coco.zip)
│   ├── element_detection/
│   │   ├── train/  val/  split3_test/  # images
│   │   ├── train.json  val.json  split3_test.json
│   │   └── train_val_chartID_list.json
│   └── plots_detection/
│       ├── bar_images/  bar_test/      # images (bar_images holds both train and val)
│       ├── bar_train.json  bar_val.json  bar_test.json
│       ├── baronly_test.json           # bar_test.json with only the `bar` class
│       └── readme.md
├── train_original/                     # original ICPR 2022 PMC training release
└── test_original/                      # original ICPR 2022 PMC test release, splits 1-5

Splits

Config	Split	COCO file	Images	Boxes
element_detection	train	element_detection/train.json	5,052	203,181
element_detection	validation	element_detection/val.json	560	22,684
element_detection	test	element_detection/split3_test.json	381	15,891
bar_detection	train	plots_detection/bar_train.json	1,635	93,139
bar_detection	validation	plots_detection/bar_val.json	182	10,048
bar_detection	test	plots_detection/bar_test.json	192	11,319

The element_detection test split is split 3 of the ICPR 2022 PMC test set, the split used for the element detection results in the paper. The bar_detection subset contains the bar charts and was used to fine-tune the CACHED bar detector.

Categories

COCO category ids start at 1. In the Parquet files, objects.category is a 0-based ClassLabel in the same order.

COCO id	Name	Description	train	val	test
1	x_title	x-axis title	3,786	437	267
2	y_title	y-axis title	4,551	505	346
3	plot_area	area enclosed by the x and y axes	5,052	560	381
4	other	other text	4,923	464	546
5	xlabel	x tick label	36,968	4,091	2,959
6	ylabel	y tick label	35,403	3,944	2,558
7	chart_title	chart title	396	38	41
8	x_tick	x tick mark	36,968	4,091	2,959
9	y_tick	y tick mark	35,403	3,944	2,558
10	legend_patch	legend marker	9,646	1,077	954
11	legend_label	legend label	9,646	1,077	954
12	legend_title	legend title	157	19	31
13	legend_area	area covering all legend items	2,850	318	238
14	mark_label	text label attached to a data mark	1,445	219	84
15	value_label	text showing a data value	5,599	726	228
16	y_axis_area	area covering the y tick marks, tick labels, and axis title	4,858	534	359
17	x_axis_area	area covering the x tick marks, tick labels, and axis title	4,982	553	371
18	tick_grouping	label that groups several ticks	548	87	57
19	bar	bar (bar_detection only)	–	–	–

The bar_detection subset has 24,741 / 2,561 / 3,059 bar boxes in train / validation / test.

Usage

With the ChartDete code (MMDetection)

The CACHED configs read the data from ./data/pmc_2022/pmc_coco/. From the root of the GitHub repo, download and unzip the COCO part (304 MB):

hf download Pengyu965/ChartDete-PMC2022 pmc_coco.zip --repo-type dataset --local-dir data/pmc_2022
unzip -oq data/pmc_2022/pmc_coco.zip -d data/pmc_2022/

Or download the full archive (2.1 GB, including the original ICPR 2022 release) and unzip it:

hf download Pengyu965/ChartDete-PMC2022 pmc_2022.zip --repo-type dataset --local-dir data
unzip -oq data/pmc_2022.zip -d data/

Then evaluate a released checkpoint from Pengyu965/ChartDete:

python tools/test.py work_dirs/cascade_rcnn_swin-t_fpn_LGF_VCE_PCE_coco_focalsmoothloss/cascade_rcnn_swin-t_fpn_LGF_VCE_PCE_coco_focalsmoothloss.py \
    work_dirs/cascade_rcnn_swin-t_fpn_LGF_VCE_PCE_coco_focalsmoothloss/checkpoint.pth --eval bbox

With 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("Pengyu965/ChartDete-PMC2022", "element_detection", split="test")  # or "bar_detection"
names = ds.features["objects"]["category"].feature.names

sample = ds[0]
image = sample["image"]  # PIL image
for bbox, cat in zip(sample["objects"]["bbox"], sample["objects"]["category"]):
    x, y, w, h = bbox  # COCO format, in pixels
    print(names[cat], [x, y, w, h])

Each row has image, image_id, file_name, width, height, and objects. objects holds parallel lists id, bbox (COCO [x, y, width, height]), category, area, iscrowd, and text. text is the transcription of a text element where the original annotation provides one, and None otherwise. image_id and objects.id match the ids in the COCO JSON files. The Parquet files were written with datasets 5.0, so use datasets>=4.0 to load them.

Bar-only annotations for the bar detector

The bar detector config (cascade_rcnn_swin-t_fpn_LGF_VCE_PCE_coco_focalsmoothloss_bar.py) trains on baronly_train.json and baronly_val.json, which are not part of the original release. They follow the same rule as the included baronly_test.json: keep only the bar annotations (category 19) and renumber them as category 1. Run this from the repo root to create them:

import json

root = "data/pmc_2022/pmc_coco/plots_detection"
for split in ["train", "val"]:
    coco = json.load(open(f"{root}/bar_{split}.json"))
    coco["annotations"] = [dict(a, category_id=1) for a in coco["annotations"] if a["category_id"] == 19]
    coco["categories"] = [{"id": 1, "name": "bar", "supercategory": None}]
    json.dump(coco, open(f"{root}/baronly_{split}.json", "w"))

Applying this rule to bar_test.json reproduces the released baronly_test.json exactly.

Source data and license

The images and the original annotations come from the PMC chart dataset of the ICPR 2022 CHART-Infographics competition, organized by Kenny Davila, Fei Xu, Saleem Ahmed, David A. Mendoza, Srirangaraj Setlur, and Venu Govindaraju. The images are taken from the Open Access subset of PubMed Central, and the PMCID of the source article is part of each file name. The competition data is licensed under CC BY-NC-SA 4.0. This dataset, including the refined annotations, is shared under the same license. Please cite the competition as well as the CACHED paper.

Citation

@inproceedings{yan2023cached,
  title     = {Context-Aware Chart Element Detection},
  author    = {Yan, Pengyu and Ahmed, Saleem and Doermann, David},
  booktitle = {Document Analysis and Recognition - ICDAR 2023},
  pages     = {218--233},
  year      = {2023},
  publisher = {Springer Nature Switzerland},
  doi       = {10.1007/978-3-031-41676-7_13}
}

@inproceedings{davila2022icpr,
  title     = {ICPR 2022: Challenge on Harvesting Raw Tables from Infographics (CHART-Infographics)},
  author    = {Davila, Kenny and Xu, Fei and Ahmed, Saleem and Mendoza, David A. and Setlur, Srirangaraj and Govindaraju, Venu},
  booktitle = {2022 26th International Conference on Pattern Recognition (ICPR)},
  pages     = {4995--5001},
  year      = {2022},
  publisher = {IEEE}
}