BurstTron

Detectron2

A computer vision project using Detectron2 for instance segmentation of type III and groups of type III SRBs on radio data (Irish-LOFAR images).

Example results

Installation

Method 1

# Clone the repository
git clone https://github.com/Hermanlrx/BurstTron.git
cd BurstTron
#This downloads and extracts the data 
curl -L -o /tmp/split_dataset.zip "https://zenodo.org/records/16759183/files/split_dataset.zip?download=1" || { echo "Dataset download failed"; exit 1; }     && unzip /tmp/split_dataset.zip -d BurstTron/     && rm /tmp/split_dataset.zip

# Create a new conda environment (Change based on your preferred env manager)
conda create -n detectron python=3.10
conda activate detectron

# Install the following in order
pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu126
pip3 install opencv-python
pip3 install 'git+https://github.com/facebookresearch/detectron2.git'

Project Structure after downloading data

├── BurstTron
|   ├── inference_script.py
|   └── training.py
|   ├── split_dataset/
|     ├── model_final.pth #Usable if user does not want to run training script
│     ├── annotations/
│     │   ├── train.json
│     │   └── val.json
│     ├── train/
│     └── val/
├── TrainingOutput/    #If your are running the training script
│   └── model_final.pth
└── README.md

Usage

Basic Inference

python inference.py --input_path /path/to/your/images

Example Commands

# Process images in a specific directory
python inference.py --input_path "/path/to/your/test_images"

Output

The script will:

1. Process all .png files in the input directory
2. Run object detection on each image
3. Generate annotated images with bounding boxes
4. Save results with filename format: originalname_event.png or originalname_no_event_detected.png
5. Create an output/ subdirectory in image folder containing all processed images

Configuration

Key Parameters for Inference.py

• SCORE_THRESH_TEST: Confidence threshold for detections (default: 0.5)
• MODEL.WEIGHTS: Path to your trained model weights
• INPUT_PATH: Directory containing images to process

Model Configuration

The script automatically loads:

• Training dataset metadata for class names
• Validation dataset for testing
• Pre-trained model weights from split_dataset/model_final.pth

Dataset Format

This project uses COCO format annotations:

{
  "images": [...],
  "annotations": [...],
  "categories": [
    {"id": 1, "name": "event_type_1"},
    {"id": 2, "name": "event_type_2"}
  ]
}

Model Training

To train your own model:

1. Donwload data from Zenodo
2. Ensure code is in the same directory as the data files
3. Run training script

Acknowledgments

• Detectron2 by Facebook AI Research
• LOFAR telescope data providers
• PyTorch and OpenCV communities

Support

For issues and questions:

• Open a GitHub issue