Computer vision
Every detector has a threshold, and somebody chose it
What the camera reports depends on a number a person picked. Which mistakes you would rather make is a business decision, not a technical one.
What a detector really returns
A vision model does not return objects. It returns guesses with a confidence attached to each one, and a threshold turns those guesses into a yes or a no.
Move the threshold up and you miss things that were really there. Move it down and you flag things that were not. There is no setting that avoids both, and pretending otherwise is how these systems lose trust in their first month.
So the threshold is set deliberately, against which mistake costs more in your situation, and it is written down with the reasoning. A missed crack in a bridge and a missed parcel on a belt are not the same kind of error.
Detection
What falls under the threshold is still there
The useful habit is to keep the low confidence detections rather than discarding them. They are the record of what the system nearly saw, and they are where the next round of improvement comes from.
They are also the honest answer when somebody asks why a system missed something. A detector that logs only its confident answers cannot explain itself, and a system that cannot explain itself is hard to keep in service.
The choices
What has to be settled before the camera is mounted
The mounting is usually the cheapest part and the one that decides the most.
| Decision | Why it comes first | What it rules out |
|---|---|---|
| Which mistake costs more | It sets the threshold and everything downstream | Any claim that the system is simply accurate |
| Lighting | It changes the data more than the model does | Sites where the light cannot be controlled or measured |
| Camera position | Resolution at the target decides what is possible | Detecting anything smaller than a few pixels |
| What happens on a miss | Somebody has to catch what falls through | Running unattended before that path exists |
Three of the four are decided with a tape measure and a light meter rather than with a model.
Or write here
Three fields, and a person reads it
Name, an address we can answer to, and what you are trying to build. You get a reference back straight away.
What do you need to see?
Tell us what has to be spotted and what it costs to miss it. That pair decides most of the build.