https://github.com/reverendbayes/yolo11m-car-damage-detector
Custom YOLO11m model for detecting and classifying car body damage (99% shattered glass, 96% flat tire detection accuracy)—optimized for high-capacity inference and assistive use in inspection and service workflows like BMW pre-loaner inspections.
https://github.com/reverendbayes/yolo11m-car-damage-detector
ai-in-automotive anomaly-detection car-inspection car-insurance cnn computer-vision convolutional-neural-networks deep-learning googlecolab image-segmentation insurance-claims object-detection oem-data-tool pretrained-model pytorch real-time-inference ultralytics yolo11 yolo11-detection
Last synced: 5 months ago
JSON representation
Custom YOLO11m model for detecting and classifying car body damage (99% shattered glass, 96% flat tire detection accuracy)—optimized for high-capacity inference and assistive use in inspection and service workflows like BMW pre-loaner inspections.
- Host: GitHub
- URL: https://github.com/reverendbayes/yolo11m-car-damage-detector
- Owner: ReverendBayes
- License: mit
- Created: 2025-04-17T23:29:38.000Z (6 months ago)
- Default Branch: main
- Last Pushed: 2025-05-01T20:42:34.000Z (5 months ago)
- Last Synced: 2025-05-04T13:17:58.235Z (5 months ago)
- Topics: ai-in-automotive, anomaly-detection, car-inspection, car-insurance, cnn, computer-vision, convolutional-neural-networks, deep-learning, googlecolab, image-segmentation, insurance-claims, object-detection, oem-data-tool, pretrained-model, pytorch, real-time-inference, ultralytics, yolo11, yolo11-detection
- Language: Jupyter Notebook
- Homepage:
- Size: 23.6 MB
- Stars: 3
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# YOLO11m Model for Car Body Damage Detection
YOLO11m-based car damage detector using deep learning, computer vision, and AI. This custom-trained model (trained.pt) was fine-tuned on a specialized dataset to detect and classify common vehicle body damage: dents, scratches, cracks, broken lamps, shattered glass, and flat tires. It’s high-capacity, fast-enough, and built for real-world inspection workflows.
---
## Model Overview
- **Architecture:** YOLO11m (231 layers, ~20M parameters)
- **Dataset:** CarDD_COCO (custom fine-tuned version)
- **Training:** Google Colab (NVIDIA L4 GPU, 22GB VRAM)
- **Classes:** dent, scratch, crack, shattered_glass, broken_lamp, flat_tire
- **Training Framework:** Ultralytics 8.3.117, PyTorch 2.6.0+cu124, Python 3.11.12
---
## Performance Snapshot
Class-by-class tactical breakdown:
| Class | Box Precision (P) | Recall (R) | mAP50 | mAP50-95 |
|:-----------------|:------------------|:-----------|:------|:---------|
| shattered_glass | 0.979 | 0.978 | 0.994 | 0.963 |
| flat_tire | 0.943 | 0.919 | 0.959 | 0.932 |
| broken_lamp | 0.826 | 0.821 | 0.895 | 0.796 |
| dent | 0.832 | 0.520 | 0.692 | 0.568 |
| scratch | 0.737 | 0.800 | 0.905 | 0.610 |
| crack | 0.699 | 0.586 | 0.620 | 0.424 |
---
## Tactical Highlights
- **Shattered Glass:** Elite — nearly flawless across all metrics.
- **Flat Tire:** Excellent — reliable, highly precise detections.
- **Broken Lamp:** Very strong — high precision and consistency.
- **Dent/Scratch/Crack:** Weakest — fair performance, but prone to missed detections and slight localization drift (especially cracks).
---
## Example Outputs
Below are sample detection results from the model:
---
## Use Cases
This model is fully usable for visual inspection support. It is intended as an assistive tool—not a replacement—for human service advisors. It aims to accelerate visual inspections, flag overlooked damage, and ensure consistency across high-throughput intake workflows. Use for inspection workflows, damage logging, or visual diagnostics.
This model saves time, adds consistency, and helps document condition clearly before keys are handed over. It’s assistive—not autonomous. It gives advisors a head start.
- Pre-loaner vehicle inspections
- Service center damage logging
- Fleet condition audits
- Insurance pre-claim imaging
- Rental return documentation
---
## Real-World Use: BMW Pre-Loaner Inspections
An enhanced version of this model is running in BMW service environments to support pre-loaner vehicle inspections.
Here’s how it works:
- Technicians capture images of the vehicle exterior.
- The model runs inference and flags visible body damage.
- Detections are mapped to a standard car diagram automatically.
- Human service advisors review and finalize the report.
This setup saves time, adds consistency, and helps document condition clearly before keys are handed over. It’s assistive—not autonomous. It gives advisors a head start.
“Good enough” here means:
- Fewer missed issues
- Fewer false alarms
- Faster workflows
- Repeatable results across locations
---
## Model Files
- `trained.pt` — trained YOLO11m checkpoint
---
## Quickstart
### 1. Install Ultralytics
```bash
pip install ultralytics
```
### 2. Load and Inference
```python
from ultralytics import YOLO
model = YOLO("trained.pt")
results = model("your_image.jpg", save=True)
results[0].show() # Visualize output
```
---
## References
- [Ultralytics GitHub](https://github.com/ultralytics/ultralytics)
---