Droid210/FleetVision

1# Model B Training Data2 3This directory contains training data for the Car Damage Detector (Model B) using Vision Transformer.4 5## Required Structure6 7```8model b/9├── train/10│   ├── Whole/          # Images of undamaged cars11│   └── Damaged/        # Images of damaged cars12└── valid/13    ├── Whole/          # Validation images of undamaged cars14    └── Damaged/        # Validation images of damaged cars15```16 17## Data Requirements18 19- **Image Format**: JPG or PNG20- **Image Size**: Any size (ViTImageProcessor will resize to 224x224)21- **Minimum Images per Set**: 10+ images (preferably 50+ for better training)22- **Class Balance**: Try to keep roughly equal numbers of Whole and Damaged images23 24## Directory Setup25 26The following directories have been created:27- `train/Whole/`28- `train/Damaged/`29- `valid/Whole/`30- `valid/Damaged/`31 32## Adding Training Data33 341. **For Whole (Undamaged) Cars**: Place images of vehicles with no visible damage352. **For Damaged Cars**: Place images of vehicles with visible damage (dents, scratches, broken parts, etc.)363. **Training/Validation Split**: Put ~80% of your data in `train/` and ~20% in `valid/`37 38## Starting Training39 40Once you have added images to the directories, run:41 42```bash43python run_damage_training.py --epochs 15 --batch-size 32 --recall-weight 2.044```45 46The trained model will be saved to: `weights/model b/best_damage_detector.pth`47 48## Demo Mode49 50For testing purposes, placeholder images have been created using sample data from Model A.51These demo images allow you to verify the training pipeline works before adding real damage detection data.52