cvtechniques/vehicle-damage-segmentation

# Vehicle Damage Instance Segmentation

## Model Description
* **Description**: This YOLOv8-seg model is designed to automate vehicle insurance claims by isolating damage areas (Dents, Scratches, Broken Glass) with pixel-level accuracy.
* **Training Approach**: Fine-tuned from a YOLOv8-seg foundation model using the Ultralytics framework.
* **Intended Use Case**: Mobile app integration to allow claimants to get immediate repair estimates, significantly reducing manual inspection wait times.

## Training Data
* **Source**: Roboflow Universe.
* **Volume**: 10,218 total images post-augmentation.
* **Classes**: Dents, Scratches, and Broken Glass.
* **Annotation Process (Original Work)**: I performed a manual audit of roughly 8 hours, refining approximately 15% of the polygon masks to ensure tighter boundaries for precise surface area calculations.
* **Split**: 70% Training, 20% Validation, 10% Testing.
* **Augmentation**: Mosaic (first 90%), Horizontal Flip, and Scale (+/- 10%).

## Training Procedure
* **Hardware**: Google Colab T4 GPU.
* **Optimizer**: AdamW | **Learning Rate**: 0.002.
* **Inference Speed**: ~3ms per frame.

## Evaluation Results
* **Overall Metrics**:
* **mAP50 (Mask)**: 0.842 (Target was 0.85).
* **Precision**: 0.864 | **Recall**: 0.771.
* **Key Findings**: Broken Glass achieved a near-perfect recall of 0.94 due to high-contrast edges.
* **Performance Analysis**: Brightness and contrast augmentations during the iteration process improved final detection accuracy by 15%.

### Key Visualizations
**Confusion Matrix**
*Shows model performance and identifies a 12% false positive rate for scratches in direct sunlight.*

**Training Results**
*Loss curves showing model convergence over the training period.*

## Visual Examples
*Representative ground truth samples showing successfully segmented damage on curved metallic panels.*

## Limitations and Biases
* **Glare**: Shiny paint reflections cause a 12% false positive rate for scratches in direct sunlight.
* **Scale**: Small scratches under 1 inch are often missed.
* **Depth**: The model provides 2D surface area but lacks 3D dent depth for volume estimation.
* **Ethical Consideration**: This model is an appraisal tool; it should not be the sole basis for final legal or financial insurance payouts without human review.