CarDD YOLOv8s — Car Damage Detection & Segmentation
YOLOv8s fine-tuned on the CarDD dataset for vision-based car damage detection and instance segmentation. The model detects and segments six types of car damage and outperforms the original paper's best model (DCN+ ResNet-101) on every individual damage class.
🗂️ Model Versions
| Version | File | Task | mAP50 |
|---|---|---|---|
| v1.0 | v1.0/best.pt |
Detection only | 71.67% |
| v2.0 | v2.0/best.pt |
Detection + Segmentation | 75.84% 🔥 |
v2.0 is the recommended version. It adds instance segmentation masks on top of detection and achieves significantly better results across all metrics.
📈 v1.0 vs v2.0 — Side by Side Comparison
Overall Metrics
| Metric | v1.0 (Detection) | v2.0 (Segmentation) | Improvement |
|---|---|---|---|
| mAP50 (Box) | 71.67% | 75.84% | ✅ +4.17% |
| mAP50-95 (Box) | 55.43% | 58.93% | ✅ +3.50% |
| Precision (Box) | 75.11% | 76.84% | ✅ +1.73% |
| Recall (Box) | 67.52% | 68.27% | ✅ +0.75% |
| Mask mAP50 | ❌ N/A | 75.90% | ✅ New |
| Mask mAP50-95 | ❌ N/A | 57.00% | ✅ New |
| Model size | 22.5 MB | 22.5 MB | ➡️ Same |
| Inference speed | 7.7ms | 61ms | ⚠️ Slower* |
*Segmentation is slower due to mask generation on top of detection.
Per-Class Comparison (Box mAP50)
| Class | v1.0 | v2.0 | Improvement |
|---|---|---|---|
| dent | 57.8% | 60.6% | ✅ +2.8% |
| scratch | 59.6% | 63.2% | ✅ +3.6% |
| crack | 37.7% | 55.5% | ✅ +17.8% |
| glass_shatter | 99.0% | 97.9% | ➡️ ~Same |
| lamp_broken | 85.5% | 86.8% | ✅ +1.3% |
| tire_flat | 90.5% | 91.0% | ✅ +0.5% |
The biggest gain is on crack (+17.8%) — the hardest class in the dataset.
📊 Full Results (v2.0)
Overall Performance (Test Set)
| Metric | Box | Mask |
|---|---|---|
| mAP@0.5 | 75.84% | 75.90% |
| mAP@0.5:0.95 | 58.93% | 57.00% |
| Precision | 76.84% | 77.40% |
| Recall | 68.27% | 68.70% |
Per-Class Results (v2.0 Test Set)
| Class | Box mAP50 | Mask mAP50 |
|---|---|---|
| glass_shatter | 97.9% | 97.9% |
| tire_flat | 91.0% | 90.2% |
| lamp_broken | 86.8% | 88.2% |
| scratch | 63.2% | 60.4% |
| dent | 60.6% | 64.9% |
| crack | 55.5% | 54.0% |
Benchmark Comparison vs CarDD Paper (v2.0)
| Model | Backbone | mAP50 |
|---|---|---|
| Mask R-CNN | ResNet-50 | 66.3% |
| Cascade Mask R-CNN | ResNet-50 | 64.7% |
| GCNet | ResNet-50 | 66.4% |
| HTC | ResNet-50 | 68.1% |
| DCN | ResNet-50 | 70.9% |
| Mask R-CNN | ResNet-101 | 67.7% |
| HTC | ResNet-101 | 68.4% |
| DCN | ResNet-101 | 69.8% |
| YOLOv8s-seg v2.0 (ours) | — | 75.8% 🔥 |
| DCN+ | ResNet-50 | 77.4% |
| DCN+ | ResNet-101 | 78.8% |
Per-Class vs DCN+ ResNet-101 (Paper's Best)
| Class | DCN+ ResNet-101 | v2.0 (Box) | Gain |
|---|---|---|---|
| dent | 40.5% | 60.6% | ✅ +20.1% |
| scratch | 34.3% | 63.2% | ✅ +28.9% |
| crack | 16.6% | 55.5% | ✅ +38.9% |
| glass_shatter | 92.6% | 97.9% | ✅ +5.3% |
| lamp_broken | 70.8% | 86.8% | ✅ +16.0% |
| tire_flat | 86.0% | 91.0% | ✅ +5.0% |
v2.0 beats DCN+ ResNet-101 on every individual damage class, while being ~10x smaller (22.5MB vs 200MB+) and requiring only a consumer GPU to train.
📉 Training Curves (v2.0)
🔢 Confusion Matrix (v2.0)
📈 PR & F1 Curves (v2.0)
🖼️ Predictions (v2.0)
📊 Dataset Label Distribution
📉 Training Curves (v1.0)
📊 Per-Class Comparison (v1.0)
🖼️ Ground Truth vs Predictions (v1.0)
🖼️ Predictions Grid (v1.0)
🚀 Usage
Install Dependencies
pip install ultralytics huggingface_hub
Load v2.0 (Recommended)
from huggingface_hub import hf_hub_download
from ultralytics import YOLO
model_path = hf_hub_download(
repo_id="abdullahg7/cardd-yolov8s",
filename="v2.0/best.pt"
)
model = YOLO(model_path)
results = model.predict(
source="your_car_image.jpg",
conf=0.25,
iou=0.45
)
results[0].show()
Load v1.0 (Detection only)
model_path = hf_hub_download(
repo_id="abdullahg7/cardd-yolov8s",
filename="v1.0/best.pt"
)
model = YOLO(model_path)
Interpret Results (v2.0)
results = model.predict("car.jpg", conf=0.25)
# Detection boxes
for box in results[0].boxes:
class_name = model.names[int(box.cls)]
confidence = float(box.conf)
print(f"Detected: {class_name} ({confidence:.2%})")
# Segmentation masks
if results[0].masks is not None:
for mask in results[0].masks.data:
print(f"Mask shape: {mask.shape}") # (H, W) binary mask
⚙️ Training Configuration (v2.0)
| Parameter | Value |
|---|---|
| Base model | yolov8s-seg.pt |
| Epochs | 50 |
| Batch size | 4 |
| Image size | 1024px |
| Optimizer | SGD |
| LR0 | 0.01 |
| LRF | 0.01 |
| Momentum | 0.937 |
| Weight decay | 0.0005 |
| Warmup epochs | 5.0 |
| Cosine LR | True |
| Close mosaic | 15 |
🔴 Damage Categories
| ID | Class | Description |
|---|---|---|
| 0 | dent | Surface deformation on car body |
| 1 | scratch | Linear paint damage |
| 2 | crack | Structural fractures |
| 3 | glass_shatter | Broken windows or windshields |
| 4 | lamp_broken | Damaged headlights or tail lights |
| 5 | tire_flat | Deflated or damaged tires |
📦 Dataset
- Name: CarDD (Car Damage Detection)
- Size: 4,000 high-resolution images, 9,000+ annotated instances
- Access: cardd-ustc.github.io (requires license agreement)
- Paper: Wang et al., IEEE T-ITS, 2023
| Split | Images | Instances |
|---|---|---|
| Train | 2,816 | 6,211 |
| Val | 810 | 1,744 |
| Test | 374 | 785 |
⚠️ Limitations
- Lower performance on hard classes (dent, scratch, crack) due to small size and visual similarity
- Trained on exterior car damage only
- Best results on clear, well-lit images
🔗 Links
- GitHub: abdullahghanem-a11y/car-damage-detection
- Dataset: cardd-ustc.github.io
📚 Citation
@article{wang2023cardd,
title = {CarDD: A New Dataset for Vision-Based Car Damage Detection},
author = {Wang, Xinkuang and Li, Wenjing and Wu, Zhongcheng},
journal = {IEEE Transactions on Intelligent Transportation Systems},
volume = {24},
number = {7},
pages = {7202--7214},
year = {2023}
}
📄 License
GNU AGPL-3.0 — in compliance with the YOLOv8 framework by Ultralytics.
⚠️ Dataset Notice: CarDD images are sourced from Flickr and Shutterstock. Proper licensing must be obtained before commercial use.