Welcome to the official repository of UniMERNet, a solution that converts images of mathematical expressions into LaTeX, suitable for a wide range of real-world scenarios.
News 🚀🚀🚀
2024.09.05 🎉🎉 Major update to the UniMERNet algorithm. This update includes enhanced algorithm details (refer to the latest version of the paper) and open-sourced training code.
2024.07.21 🎉🎉 Add Math Formula Detection (MFD) Tutorial based on PDF-Extract-Kit MFD model.
2024.06.06 🎉🎉 Open-sourced evaluation code for UniMER dataset.
2024.05.06 🎉🎉 Open-sourced UniMER dataset, including UniMER-1M for model training and UniMER-Test for MER evaluation.
2024.05.06 🎉🎉 Add Streamlit formula recognition demo and provided local deployment App.
2024.04.24 🎉🎉 Paper now available on ArXiv.
2024.04.24 🎉🎉 Inference code and checkpoints have been released.
Demo Video
https://github.com/opendatalab/UniMERNet/assets/69186975/ac54c6b9-442c-48b0-95f9-a4a3fce8780b
https://github.com/opendatalab/UniMERNet/assets/69186975/09b71c55-c58a-4792-afc1-d5774880ccf8
Quick Start
Clone the repo and download the model
git clone https://github.com/opendatalab/UniMERNet.git
cd UniMERNet/models
git lfs install
git clone https://huggingface.co/wanderkid/unimernet
git clone https://www.modelscope.cn/wanderkid/UniMERNet.git
Installation
Create a clean Conda environment
conda create -n unimernet python=3.10
conda activate unimernet
Method 1: Install via pip (recommended for general users)
pip install --upgrade unimernet
pip install "unimernet[full]"
Method 2: Local installation (recommended for developers)
pip install -e ."[full]"
Running UniMERNet
-
Streamlit Application: For an interactive and user-friendly experience, use our Streamlit-based GUI. This application allows real-time formula recognition and rendering.
unimernet_gui
Ensure you have the latest version of UniMERNet installed (pip install --upgrade unimernet & pip install "unimernet[full]"
) for the streamlit GUI application.
-
Command-line Demo: Predict LaTeX code from an image.
python demo.py
-
Jupyter Notebook Demo: Recognize and render formula from an image.
jupyter-lab ./demo.ipynb
Performance Comparison (BLEU) with SOTA Methods.
UniMERNet significantly outperforms mainstream models in recognizing real-world mathematical expressions, demonstrating superior performance across Simple Printed Expressions (SPE), Complex Printed Expressions (CPE), Screen-Captured Expressions (SCE), and Handwritten Expressions (HWE), as evidenced by the comparative BLEU Score evaluation.
![BLEU](./asset/papers/fig1_bleu.jpg)
Visualization Result with Different Methods.
UniMERNet excels in visual recognition of challenging samples, outperforming other methods.
![Visualization](https://github.com/opendatalab/VIGC/assets/69186975/6edcac69-5082-43a2-8095-5681b7a707b9)
UniMER Dataset
Introduction
The UniMER dataset is a specialized collection curated to advance the field of Mathematical Expression Recognition (MER). It encompasses the comprehensive UniMER-1M training set, featuring over one million instances that represent a diverse and intricate range of mathematical expressions, coupled with the UniMER Test Set, meticulously designed to benchmark MER models against real-world scenarios. The dataset details are as follows:
UniMER-1M Training Set:
- Total Samples: 1,061,791 Latex-Image pairs
- Composition: A balanced mix of concise and complex, extended formula expressions
- Aim: To train robust, high-accuracy MER models, enhancing recognition precision and generalization
UniMER Test Set:
- Total Samples: 23,757, categorized into four types of expressions:
- Simple Printed Expressions (SPE): 6,762 samples
- Complex Printed Expressions (CPE): 5,921 samples
- Screen Capture Expressions (SCE): 4,742 samples
- Handwritten Expressions (HWE): 6,332 samples
- Purpose: To provide a thorough evaluation of MER models across a spectrum of real-world conditions
Dataset Download
You can download the dataset from OpenDataLab (recommended for users in China) or HuggingFace.
Download UniMER-Test Dataset
Download the UniMER-1M dataset and extract it to the following directory:
./data/UniMER-1M
Download the UniMER-Test dataset and extract it to the following directory:
./data/UniMER-Test
Training
To train the UniMERNet model, follow these steps:
-
Specify the Training Dataset Path: Open the configs/train
fold and set the path to your training dataset.
-
Run the Training Script: Execute the following command to start the training process.
bash script/train.sh
Notes:
- Ensure that the dataset path specified in the
configs/train
fold is correct and accessible. - Monitor the training process for any errors or issues.
Testing
To test the UniMERNet model, follow these steps:
-
Specify the Test Dataset Path: Open the configs/val
fold and set the path to your test dataset.
-
Run the Test Script: Execute the following command to start the testing process.
bash script/test.sh
Notes:
- Ensure that the dataset path specified in the
configs/val
fold is correct and accessible. - The
test.py
script will use the specified test dataset for evaluation. Remember to change the test set path in test.py to your actual path. - Review the test results for performance metrics and any potential issues.
Math Formula Detection Tutorial
The prerequisite for formula recognition is to detect the areas within PDF or webpage screenshots where formulas are located. The PDF-Extract-Kit includes a powerful model for detecting formulas. If you wish to perform both formula detection and recognition by yourself, you can refer to the Formula Detection Tutorial for guidance on deploying and using the formula detection model.
TODO
Citation
If you find our models / code / papers useful in your research, please consider giving us a star ⭐ and citing our work 📝, thank you :)
@misc{wang2024unimernet,
title={UniMERNet: A Universal Network for Real-World Mathematical Expression Recognition},
author={Bin Wang and Zhuangcheng Gu and Chao Xu and Bo Zhang and Botian Shi and Conghui He},
year={2024},
eprint={2404.15254},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
Acknowledgements
- VIGC. The model framework is dependent on VIGC.
- Texify. A mainstream MER algorithm, UniMERNet data processing refers to Texify.
- Latex-OCR. Another mainstream MER algorithm.
- Donut. The UniMERNet's Transformer Encoder-Decoder are referenced from Donut.
- Nougat. The tokenizer uses Nougat.
Contact Us
If you have any questions, comments, or suggestions, please do not hesitate to contact us at wangbin@pjlab.org.cn.
License
Apache License 2.0