RF-DETR Roof Damage Detection

This project uses the 'RFDETR_base' model from Roboflow to train a vision model capable of detecting hail damage and wind damage on roofs.

---

Installation

Clone the repository

git clone https://github.com/LifeLiveOn/Modification-to-RF-DETR-model

cd Modification-to-RF-DETR-model

Install uv

On Windows: pip install uv

On Linux: curl -LsSf https://astral.sh/uv/install.sh | sh

Set up a Python 3.10 virtual environment

uv venv --python 3.10

Install dependencies

uv pip install -r requirements.txt

Install local RF-DETR modifications

• cd rf-detr-modifications
• uv pip install -e .

---

Running the Demo

From the project root, launch the Streamlit demo: streamlit run app.py

---

Utilities Overview (utils/)

Script: check_miss_path.py

Description: Ensures that all images referenced in 'annotation.json' exist in the dataset directories.

Script: merge_coco.py

Description: Merges multiple COCO-format JSON files into one unified file, preserving source references.

Script: remap_label.py

Description: Remaps category IDs to maintain consistent 'category_id' values across annotation files.

Script: tiles_images.py

Description: Preprocesses images into tiles of specified sizes for training purposes.

---

Main CLI Script

main.py is used for training, testing, and inference through the command line.

Example Usage

Inference (normal mode): python main.py --mode test --infer_mode normal --tile_size small

Inference (tiled mode): python main.py --mode test --infer_mode tiled --tile_size small --path datasets/hail_1/test

Training: python main.py --mode train

Test on wind damage dataset: python main.py --mode test --infer_mode normal --path datasets/wind_1/test

---

Exporting to ONNX

Note: ONNX export only works with Python 3.10 due to 'onnxsim' incompatibility with Python 3.12.

Install ONNX tools:

uv pip install onnx onnxsim onnxscript

Export the model:

python export_to_onnx.py

---

Project Summary

• Purpose: Detect hail and wind roof damage using the RFDETR model.
• Framework: Based on RFDETR (DETR variant) with local modifications.
• Training Pipeline: Supports normal and tiled inference, data preprocessing utilities, and ONNX export.
• Environment: Requires Python 3.10 for full compatibility.

---

Repository Structure

Modification-to-RF-DETR-model/ ├── app.py ├── main.py ├── requirements.txt ├── rf-detr-modifications/ ├── utils/ │ ├── check_miss_path.py │ ├── merge_coco.py │ ├── remap_label.py │ └── tiles_images.py └── datasets/ ├── hail_1/ └── wind_1/

---

Author: LifeLiveOn
Model Base: RFDETR (Roboflow)
License: MIT