Synthetic Images for AI Training
Synthetic highway images, rendered in 3D and labelled automatically frame by frame, to train a computer-vision AI for an R&D project with an Italian motorway operator.
The brief
An R&D project for an Italian motorway operator needed to train a computer-vision AI. An AI like this learns from images, a great many of them, each one labelled with what the model should recognise. Collecting and labelling that footage on a live motorway is slow, expensive and, for the situations that matter most, often unsafe.
What we did
We built a virtual stretch of motorway and generated the training images in CGI: lanes, guardrails, gantries and traffic of cars, vans and trucks, with changing positions, light and shadows. Every frame comes out already labelled, with boxes marking the areas and objects the AI has to learn, so there is no manual tagging.
Why synthetic data
- Perfect labels. The scene is ours, so every annotation is exact and free.
- Any situation, on demand. Rare or risky scenarios can be rendered as often as needed, without anyone setting foot on the carriageway.
- Endless variation. Traffic, light and camera can change from one image to the next, which helps the AI generalise.
- No privacy issues. No real number plates or people.
This was one of our earliest AI projects, made in 2022, before AI image generators went mainstream: here AI was the student, and CGI was the teacher.
More on our AI-assisted production.





