Hail Damage Detection — YOLO Pipeline

YOLOv11 training pipeline for detecting hail and wind damage from drone footage, built to compare against the existing RF-DETR implementation.

Project Structure

hail-yolo/
  annotations/
    train_annotations.coco.json
    valid_annotations.coco.json
    test_annotations.coco.json
  datasets/
    train/
      images/        ← place training images here
      labels/        ← auto-generated by preprocess.py
    valid/
      images/        ← place validation images here
      labels/
    test/
      images/        ← place test images here
      labels/
  preprocess.py      ← converts COCO annotations → YOLO .txt labels
  train.py           ← trains YOLOv11 on the dataset
  data.yaml          ← dataset config for YOLO

Requirements

• Python 3.10+
• ultralytics

Install dependencies:

python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install ultralytics

How to Run

1. Add images

Drop the dataset images into their corresponding folders:

datasets/train/images/
datasets/valid/images/
datasets/test/images/

Image filenames must match those referenced in the COCO JSON files.

2. Run preprocessing

Converts the 3 COCO JSON annotation files into YOLO-compatible .txt label files:

python3 preprocess.py

Labels will be generated in each split's labels/ folder automatically.

3. Train the model

python3 train.py

Results are saved to runs/hail_yolo_v1/.

Classes

ID	Name
0	wind
1	hail

Notes

• All images are 640×640 (pre-tiled from original drone footage)
• Annotations sourced from the RF-DETR repo: LifeLiveOn/Modification-to-RF-DETR-model
• This pipeline is intended to benchmark YOLOv11 against the existing RF-DETR implementation