The challenge
In autonomous driving, the gap between a safe model and an unsafe one is often the quality of its training data. Footage from vehicle-mounted cameras is cluttered: objects overlap, sit at different distances and appear under changing light.
The client needed pixel-precise boxes on objects near and far, object IDs that stayed consistent from frame to frame so the model could learn motion, and several classes labelled at once.
What we did
Multi-class labelling. We labelled four classes critical to city driving: vehicles (cars, trucks, motorcycles, buses), pedestrians, traffic lights including their signal state, and licence plates.
Video tracking. Using CVAT's tracking and interpolation, each moving object kept the same ID across frames, giving the model the data it needs to understand speed and direction.
Two-pass QA. Senior annotators reviewed every task before export, with particular care on small, distant and partly hidden objects.
The result
A clean, consistently labelled dataset that let the client train faster and improve detection of pedestrians and traffic signals in low-light streets — plus a repeatable process for labelling thousands of hours of road footage at the same quality.
