From high-resolution gate-camera images we detected and localised vehicle damage as oriented bounding boxes — rotated rectangles fitted to the damage at any angle — and classified each into a fine-grained set of damage types.
The challenge
Gate cameras see vehicles under varying light and angles, and damage appears at arbitrary orientations: diagonal scratches, bent panels, cracked lights. Standard axis-aligned boxes fit such cases poorly, so the client needed orientation-aware detection.
Our approach
We used an oriented object-detection model that detects, localises (with rotated boxes) and classifies each instance into one of 11 damage classes spanning body, glass, lights, tyres and rims. Every output carries its damage class, confidence, box parameters (centre, width, height, angle) and, optionally, the associated vehicle part.
Results & benefits
- 11 fine-grained damage classes across body, glass, lights, tyres and rims.
- Oriented boxes that fit rotated damage precisely.
- A drop-in for automated inspection — rental returns, pre-purchase checks, insurance claims and fleet maintenance.
- Pairs with part segmentation to map each damage to a specific part.
How it was built
The detector is an oriented-bounding-box model (YOLO-OBB / Rotated R-CNN class) tuned to the gate-camera viewpoint.
