From high-resolution gate-camera images we performed fine-grained segmentation of vehicle parts, returning each panel as a polygon attributed with its part class and direction (left/right, front/rear) — the foundation for automated vehicle inspection.
The challenge
A gate camera gives a controlled viewpoint, but lighting, reflections, dirt and accessories, and the sheer variety of vehicle shapes (sedan, SUV, truck, hatchback) make part-level segmentation hard — and telling a left part from its right-side twin is harder still.
Our approach
We used instance-segmentation models that isolate the vehicle from its background and then segment it into its constituent parts with direction-aware labels:
- each part detected as a pixel-precise polygon;
- a direction attribute — left, right, front, rear or center — assigned to every part;
- output attributes including the base part class, side, confidence and pixel area (with real-world area where the camera is calibrated).
Results & benefits
- Pixel-precise, direction-aware part polygons from a real operational camera feed.
- Directional inspection — left-side vs right-side damage, a missing right mirror, a failed left headlight.
- A clean base layer to attach damage detections to specific parts.
