Extract Line Chart Data
A repo that shows how to automatically extract the data of a line chart. Mainly a wrapper around LineFormer and ChartDete.
Other Solutions
There's other solutions out there:
- PP-Chart2Table (Huggingface) - haven't tried and would love to hear opinions!
- Graph2Table - commercial SaaS, superior extraction that plextract but only 3 plots per day for free (as of 13.11.25)
- Deplot - didn't work well for me on the example plots
- Matcha ChartQA
- Mostly the mentioned models as a collection on Huggingface
Installation
Clone the repository and follow the steps below for either Modal or local execution.
Modal (cloud, recommended)
git clone https://github.com/tdsone/extract-line-chart-data.git cd extract-line-chart-data # install the modal extra locally uv pip install -e ".[modal]"
You need a modal.com account and the Modal CLI configured.
Local execution (GPU required)
git clone https://github.com/tdsone/extract-line-chart-data.git cd extract-line-chart-data # create / activate a Python 3.10 virtualenv (example with uv) uv venv --python 3.10 source .venv/bin/activate # install project dependencies uv pip install -e ".[local]" # pull third-party models (mmcv-full + custom ChartDete fork) bash setup_local_env.sh
setup_local_env.sh runs mim install mmcv-full, clones ChartDete into third_party/ChartDete, and installs its mmdet fork with pip install --no-build-isolation -e.
Usage
All images in the folder input will be processed and results saved to output.
Python API
from plextract import extract # Run locally (requires plextract[local]) extract(input_dir="input", output_dir="output", backend="local") # Or run on Modal cloud (requires plextract[modal]) extract(input_dir="input", output_dir="output", backend="modal")
Input/Output Folder Structure
With Modal:
<run_id>/
├── input
│ ├── input1.jpeg
│ ├── input2.jpeg
│ └── input3.png
└── output
├── input1.jpeg
│ ├── axis_label_texts.json # Text extracted from axis labels
│ ├── chartdete
│ │ ├── bounding_boxes.json
│ │ ├── cropped_xlabels_0.jpg # Cropped images of axis labels
│ │ ├── ...
│ │ ├── cropped_ylabels_0.jpg
│ │ ├── ...
│ │ ├── label_coordinates.json # Coordinates of the detected elements
│ │ └── predictions.jpg # Image with bounding boxes of detected elements
│ ├── converted_datapoints
│ │ ├── data.json # The extracted data!
│ │ └── plot.png # The plot generated from the extracted data
│ └── lineformer
│ ├── coordinates.json # The image relative coordinates of the lines
│ └── prediction.png
├── input2.jpeg
│ ├── ...
└── input3.png
├── ...
14 directories, 60 files
How It Works
The pipeline works as follows:
- Use ChartDete to detect chart elements, most importantly axis labels and the plot area.
- OCR the numbers from the labels.
- Extract the coordinates of the lines in the line chart using LineFormer.
- Correct the coordinates of the lines to be relative to the plot origin.
- Calculate the conversion from pixels to axis values.
- Convert the coordinates using the conversion parameter from step before.
Example
Input
Output
This chart was generated using matplotlib using the extracted data (examples/input2.png/data.json)
Resources
Contact
If you need help setting this up or would just like to use it, shoot me an email: mail@timonschneider.de