Every vision demo looks the same. A part slides past, a green box appears, a number goes up. It is convincing and it tells you almost nothing, because the demo was built on the vendor’s parts under the vendor’s lights.
The useful question is narrower: on your line, on your parts, what can a camera decide more reliably than the person currently deciding it?
Where a camera is genuinely better
Three things, and they are not the ones people expect.
Consistency, not accuracy. A good inspector on a good day beats a camera. The camera’s advantage is that it has no bad days, no end of shift, and no eleventh hour of a twelve-hour rotation. If your scrap rate has a shape that follows the roster, that gap is what a camera closes.
Everything, not a sample. Manual inspection is almost always sampling — every tenth part, or every part on a line slow enough to allow it. A camera inspects all of them at line speed. That changes the question you can answer afterwards from “was this batch acceptable” to “which units in this batch were not”.
Memory. A person who rejects a part remembers it for about a shift. A camera keeps the image, the measurement and the verdict, tied to batch, machine and time. When a customer calls about one unit eight months later, that is the difference between an apology and an answer.
Where it is worse, and stays worse
Anything requiring context. “This scratch doesn’t matter because it’s on the face that gets covered” is a judgement built on knowing the product. That reasoning can be encoded, but each rule costs money and every one of them is a thing to maintain.
Defects you cannot show it. A model learns from examples. A failure mode that occurs twice a year will not be in the training set, and the camera will pass it confidently. This is the honest limit of the technology and the one most often glossed over: vision finds the defects you already know about, faster and more consistently. It does not discover new ones.
Anything where the lighting moves. Sunlight through a roof light, a reflective surface, a part whose orientation is not fixed. All solvable, none free. Most vision projects that disappoint were sold as software and were actually a lighting and fixturing problem wearing a software price tag.
What it costs to find out
Not much, and the sequence matters.
- Photographs first. A few hundred images of good parts and, crucially, of the defects you actually care about. If you cannot produce examples of the defect, that is the finding — and it is better to learn it in week one.
- An honest accuracy number, before hardware. From those images you can estimate what a model will achieve. If the answer is 92% and you need 99.9%, stop there. That conversation costs a week.
- One station, one defect class. Not the whole line. The narrowest useful problem, running in production, measured against what the bench currently catches.
- Then widen — more classes, more stations — on evidence rather than on a roadmap.
The number that decides it
Take your annual cost of the defect escaping: returns, rework, credit notes, the customer conversation. Then your annual cost of catching it: inspector hours plus the parts scrapped unnecessarily because a tired judgement was conservative.
If those two numbers together are small, no camera will pay for itself, and anyone who tells you otherwise is selling. If they are large and one of them is mostly labour spent looking, that is exactly the shape a camera fits.
We would rather tell you which one you have in the first call than after a purchase order. Send a handful of images — good parts and the defects that bother you — and we will give you a straight read.