README.md

---
license: mit
task_categories:
- object-detection
- image-segmentation
language:
- en
tags:
- chemistry
- reaction
- ocr
- molecular-structure
---

# RxnLabelData

Annotated chemical reaction diagram dataset for the paper: *RobustRDP: Advancing Reaction Diagram Parsing via Synthetic-to-Real Data Scaling and Robustness-Oriented Training*.

## Description

This dataset contains 3,500 annotated chemical reaction diagram images with bounding box and reaction relationship annotations. The annotations were created using the [RxnLabel](https://github.com/jaydetang/RxnLabel) annotation platform.

### Annotation Workflow

1. **Object Detection**: A YOLO model automatically detects molecules (Mol) and text (Txt) regions in reaction images.
2. **Relationship Annotation**: An interactive web-based tool is used to label relationships between detected objects, grouping them into reactants, conditions, and products.

## Data Structure

```
dataset/
├── images.zip       # ZIP archive containing all PNG images (3,500)
├── labels.zip       # ZIP archive containing all JSON annotation files (3,500)
└── README.md
```

## Quick Start

Download and extract the archives:

```bash
# Download from Hugging Face
# Option 1: Using git
git clone https://huggingface.co/datasets/Jingcz/RxnLabelData
cd RxnLabelData
unzip images.zip
unzip labels.zip

# Option 2: Using Python
from huggingface_hub import snapshot_download
snapshot_download(repo_id="Jingcz/RxnLabelData", local_dir="./RxnLabelData")
```

After extraction, you will have:

```
RxnLabelData/
├── images/          # 3,500 PNG images (000000.png ~ 003499.png)
├── labels/          # 3,500 JSON annotation files (000000.json ~ 003499.json)
├── images.zip
├── labels.zip
└── README.md
```

## Label Format

Each JSON file contains:

```json
{
  "id": 0,
  "width": 1496,
  "height": 1370,
  "file_name": "000000.png",
  "bboxes": [
    {
      "id": 0,
      "bbox": [66.88, 183.59, 37.72, 51.03],
      "category_id": 3
    }
  ],
  "reactions": [
    {
      "reactants": [0],
      "conditions": [1, 2],
      "products": [3]
    }
  ]
}
```

### Category IDs

| ID | Label | Description |
|----|-------|-------------|
| 1 | Mol | Molecule structure |
| 2 | Txt | Text / condition label |
| 3 | Idt | Identifier / arrow label |

## Usage Example

```python
import json
from PIL import Image

# Load image
img = Image.open("images/000000.png")

# Load annotations
with open("labels/000000.json") as f:
    data = json.load(f)

# Access bounding boxes
for box in data["bboxes"]:
    print(f"ID: {box['id']}, Category: {box['category_id']}, Box: {box['bbox']}")

# Access reactions
for rxn in data["reactions"]:
    print(f"Reactants: {rxn['reactants']}")
    print(f"Conditions: {rxn['conditions']}")
    print(f"Products: {rxn['products']}")
```

## Citation

If you use this dataset, please cite:

```bibtex
@article{robustrdp,
  title={RobustRDP: Advancing Reaction Diagram Parsing via Synthetic-to-Real Data Scaling and Robustness-Oriented Training},
  author={...},
  year={2025}
}
```

## Related Resources

- **Annotation Platform**: [RxnLabel](https://github.com/jaydetang/RxnLabel) - Web-based tool for annotating chemical reaction diagrams