🚗 Car Damage Detection

A deep learning project for detecting and classifying car damage using computer vision. The project covers multiple tasks — classification, object detection, and instance segmentation.

---

📌 Project Overview

Given an image of a damaged car, the models can:

• Classify what types of damage are present (multi-label classification)
• Locate exactly where the damage is (object detection with bounding boxes)
• Outline the exact shape of each damage (instance segmentation with masks)

---

🗂️ Repository Structure

car-damage/
├── classification/
│   └── car_damage_classification.ipynb       # Multi-label classification pipeline
├── object_detection/
│   └── car_damage_detection.ipynb            # YOLOv8 object detection pipeline
├── instance_segmentation/
│   └── car_damage_segmentation.ipynb         # YOLOv8 instance segmentation pipeline
├── .gitignore
├── LICENSE
└── README.md

---

📊 Dataset

• Source: CarDD — Car Damage Detection Dataset on Kaggle
• Size: ~4,000 high-resolution images, 9,000+ damage instances
• Format: COCO JSON annotations (bounding boxes + segmentation masks)
• Split: Train (2,816) / Val (810) / Test (374)

🏷️ Damage Categories (6 Classes)

Class	# Images
Scratch	2,121
Dent	1,751
Lamp Broken	693
Glass Shatter	674
Crack	604
Tire Flat	309

---

📁 Task 1 — Multi-Label Classification

Model

• Backbone: EfficientNet-B3 pretrained on ImageNet
• Strategy: Transfer learning — frozen early layers, fine-tuned last 3 blocks
• Head: Custom classifier (Linear → ReLU → Dropout → Linear)
• Output: 6 independent probabilities (one per damage class)

Training

Parameter	Value
Epochs	60
Batch size	32
Optimizer	AdamW
Loss	BCEWithLogitsLoss (weighted)
Scheduler	CosineAnnealingLR
Image size	224 × 224
Device	GPU (CUDA)

Results (Test Set)

Class	ROC-AUC	F1 @ 0.5	F1 @ Best Threshold
Glass Shatter	0.9899	0.8535	0.8986
Tire Flat	0.9880	0.6988	0.8519
Lamp Broken	0.8994	0.5846	0.6567
Dent	0.8502	0.7411	0.7514
Scratch	0.8428	0.7838	0.7911
Crack	0.8074	0.4062	0.4912
Overall	0.8963	0.6789	0.7568

📝 Per-class thresholds were tuned using Precision-Recall curves on the test set.

---

📁 Task 2 — Object Detection

Model

• Architecture: YOLOv8m pretrained on COCO
• Strategy: Fine-tuned on CarDD dataset
• Input: COCO annotations converted to YOLO format
• Output: Bounding boxes + class labels + confidence scores

Training

Parameter	Value
Epochs	100
Batch size	16
Image size	640 × 640
Device	GPU (CUDA)
Early stopping	patience = 15

Results (Test Set)

Class	mAP50	mAP50-95
Glass Shatter	0.986	0.937
Tire Flat	0.936	0.902
Lamp Broken	0.889	0.781
Dent	0.618	0.373
Scratch	0.585	0.336
Crack	0.499	0.262
Overall	0.752	0.599

---

📁 Task 3 — Instance Segmentation

Model

• Architecture: YOLOv8m-seg pretrained on COCO
• Strategy: Fine-tuned on CarDD dataset
• Input: COCO polygon annotations converted to YOLO segmentation format
• Output: Bounding boxes + pixel-level masks per damage instance

Training

Parameter	Value
Epochs	100
Batch size	16
Image size	640 × 640
Device	GPU (CUDA)
Early stopping	patience = 15

Results (Test Set)

Class	Box mAP50	Box mAP50-95	Mask mAP50	Mask mAP50-95
Glass Shatter	0.991	0.940	0.991	0.920
Tire Flat	0.895	0.884	0.895	0.887
Lamp Broken	0.885	0.778	0.885	0.769
Dent	0.631	0.387	0.642	0.362
Scratch	0.611	0.360	0.598	0.297
Crack	0.589	0.349	0.566	0.237
Overall	0.767	0.617	0.763	0.578

---

📈 All Tasks Comparison

Class	Classification F1	Detection mAP50	Segmentation mAP50
Glass Shatter	0.90	0.986	0.991
Tire Flat	0.85	0.936	0.895
Lamp Broken	0.66	0.889	0.885
Dent	0.75	0.618	0.642
Scratch	0.79	0.585	0.598
Crack	0.49	0.499	0.566
Overall	0.757	0.752	0.763

---

🚀 How to Run

1. Clone the repo:

git clone https://github.com/Elsaraf1/car-damage.git
cd car-damage

2. Download the dataset from Kaggle and place it in the root folder.

3. Open the notebook for the task you want:

   • classification/car_damage_classification.ipynb
   • object_detection/car_damage_detection.ipynb
   • instance_segmentation/car_damage_segmentation.ipynb

4. For best results run on Kaggle with GPU:

   • Enable GPU: Settings → Accelerator → GPU T4 x2
   • Add the dataset directly from Kaggle

---

🔜 Roadmap

• Multi-label classification (EfficientNet-B3)
• Object detection (YOLOv8m)
• Instance segmentation (YOLOv8m-seg)
• Salient object detection

---

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.