Detection is a latency budget

In a notebook, a computer vision model is judged by its mAP score. On a road, it's judged by a stopwatch. The gap between those two judgments is where most real-time vision projects quietly fail.

Working on road-safety detection taught me to treat latency as a budget with line items: capture, preprocessing, inference, transport, decision. Every line item wants more than its share. The model wants a bigger input resolution. The network wants a retry. The pipeline wants one more filter pass.

The discipline is deciding, before training anything, what the total budget is — the time between a hazard appearing and an alert being actionable — and then making every component justify its slice. A YOLO variant that scores two points lower but runs twice as fast is not a compromise; at road speed, it's the better model.

5G changes the shape of the budget, not the existence of it. Lower transport latency doesn't excuse a slow pipeline — it exposes it. When the network stops being your bottleneck, your preprocessing becomes the thing everyone stares at.

So now every vision project I start begins with the same question, and it isn't 'which model?' It's: how many milliseconds do we actually have — and who's spending them?