What 96%+ accuracy means, and how we measure it.
An accuracy figure is only useful if you know where it came from. Here is where ours comes from, what changes it, and how to test it on your own footage before you pay for anything.
How it is measured
We compared our AI's counts with manual counts of the same footage, on custom projects for clients. The counts agreed to 96% or better. In plain terms: where people counted 1,000 vehicles, Wayro counted between 960 and 1,040.
It is a measured figure from real projects, not a promise for every camera. An elevated view of a clear road does better. A low camera looking across dense traffic at night does worse. The only way to know for your camera is to test it, and that is free.
The engineer check
Most wrong counts do not come from the AI missing a vehicle. They come from a line or an area in the wrong place: a crossing line drawn where traffic queues, an entry area behind a sign, two areas that overlap.
So every project is checked by a trained engineer before the AI processes it. Once the invoice is paid, the engineer goes through every marking on every camera and corrects anything that would count wrongly. It is free on every project, whatever its size, and takes 24 to 72 hours depending on how much footage there is.
What makes accuracy better or worse
- Angle and height. An elevated view along the road is better than a low view across it. Vehicles should be separate, not hidden behind one another.
- Things in the way. Poles, signs, gantries and large vehicles hiding smaller ones are the usual reason a vehicle is not counted. Place lines and areas where the view is clear.
- Light and weather. Night, heavy rain and glare are harder than daylight.
- Resolution and frame rate. Any resolution works. More pixels on each vehicle and a steady frame rate help at a distance.
Check it yourself
- Open the real data. Each example on our home page has its real data to download as CSV, Excel or JSON. Play the clip and compare.
- Use the free hours on your most difficult camera. Your first project includes 12 free hours. Send the camera you trust least, at the time of day you trust least.
- Count a sample by hand. Pick 15 minutes of that footage, count it yourself, and compare your totals with ours, class by class.
- Watch the annotated video. It costs nothing extra: your footage with a box and a label drawn on every vehicle we followed, so any missed vehicle is easy to find.
If the numbers are not right for your camera, you have paid nothing, and you know before you commit.