Overview

Developing a YOLOv8-based computer vision system to detect and localize vehicle damage from images. The model identifies damaged regions and classifies damage types, enabling automated inspection workflows.

Goal: Reduce manual inspection effort and enable faster, more consistent vehicle damage assessment for real-world use cases such as insurance claims.

Status: Model training and evaluation completed; currently working on deployment and user interface.

Dataset

• Annotated dataset with bounding boxes for multiple vehicle damage types
• Includes real-world variations (lighting, angles, backgrounds)
• Preprocessed and structured for object detection training
• Split into training and validation sets

Approach

• Trained YOLOv8 model for damage detection and localization
• Applied data augmentation to improve generalization
• Optimized training for balanced precision and recall
• Generated predictions with bounding boxes and confidence scores
• Evaluated performance using standard detection metrics

Model

YOLOv8 (You Only Look Once) is used for efficient, real-time object detection. The model processes images in a single pass, predicting both damage location and type, making it suitable for practical deployment scenarios.

Results

• Precision: 0.69
• Recall: 0.67
• mAP@0.5: 0.68
• mAP@0.5:0.95: 0.54
• Achieves reliable detection performance across varied real-world conditions

Key Insights

• Model performance is highly sensitive to image quality and lighting
• Diverse and well-annotated data significantly improves detection accuracy
• YOLO enables efficient end-to-end detection in a single pipeline

Next Steps

• Develop a Streamlit-based interface for real-time predictions
• Export structured outputs (bounding boxes, labels, confidence)
• Deploy model as an API for integration into applications
• Extend system with damage severity estimation