What a camera can and cannot see
A useful discipline before any computer vision project: write down what the camera physically cannot observe, and check whether the thing you actually care about is on that list.
It surprisingly often is.
A client wanted to detect unsafe lifting on a shop floor. The model could see posture. It could not see the weight of the object, which is most of what makes a lift unsafe. We could have shipped something with good accuracy against the labels and no relationship to the outcome.
Three questions worth asking early
- Is the signal visible at all? Temperature, weight, torque and intent are not. Sometimes a proxy is good enough; you should know you’re using one.
- Is it visible from where the camera is? Occlusion kills more deployments than model quality does.
- Would a person watching this footage get it right? If a trained human can’t, a model won’t, and your labels will be noise.
The third is the cheap one. Before commissioning anything, have someone sit with an hour of recorded footage and try to do the task by eye. It costs an afternoon and it has killed more bad projects than any amount of model evaluation.
The part nobody budgets for
Assume the model works. Now: who sees the alert, how fast, and what are they empowered to do about it? An unsafe-lift detector that emails a weekly summary to a safety manager is a reporting tool, not a safety intervention. That’s a legitimate thing to build — but it should be a decision, not something you discover after go-live.
The camera is the easy part. It was always going to be the easy part.