# ABot-OCR

ABot-OCR is a document image OCR model that converts PDF/document page images into structured **Markdown** output, supporting recognition and reconstruction of text, mathematical formulas (LaTeX), tables (HTML), and other elements.

Code: https://github.com/amap-cvlab/ABot-OCR

Paper: https://arxiv.org/abs/2605.27978 

## Benchmarks

![ABot-OCR Benchmark Results](./metric.png)


## Requirements

Python 3.11 is recommended. Install the following dependencies:

```bash
pip install vllm==0.18.0 torch==2.10.0
```

> **Note:** Inference uses vLLM to load the model. Sufficient GPU memory is required (~4GB model weights; actual usage depends on `batch_size` and image resolution).

---

## Inference

Inference script: [`abot-ocr-infer.py`](./abot-ocr-infer.py)

### 1. Configure Model Path

Update the default model path in the script:

```python
MODEL_PATH = "./abot-ocr"  # Path to the model directory in this repo
```

### 2. Run from Command Line

Edit the parameters in the `__main__` block at the bottom of `abot-ocr-infer.py`, then run:

```bash
python abot-ocr-infer.py
```

---

## Acknowledgements

Our work is inspired by many excellent open-source projects. We sincerely thank the developers of [Qwen-VL](https://github.com/QwenLM/Qwen-VL), [PaddleOCR-VL](https://github.com/PaddlePaddle/PaddleOCR), [MinerU](https://github.com/opendatalab/MinerU), and the broader OCR community.