In computer vision, a better light beats a better model

Computer vision is about as concrete as AI gets. A camera sees something, the system works out what it is, and a person or a workflow does something about it. In regional industry that lands on real jobs: inspections, counting stock, safety checks, catching damaged product, watching an asset, capturing evidence before a truck leaves the yard.

And here’s the thing most vision projects get wrong. People obsess over the model and barely think about the camera, when it’s almost always the camera and the conditions that decide whether the whole thing works. A better light, a fixed mounting point and a cleaner angle will do more for accuracy than swapping in a fancier model, and cost a fraction of it. The model is the easy part now. Getting a machine to reliably see the thing, in a dusty shed with glare off a roller door and the subject moving, is the actual project.

Test the capture before you spend on the clever bit

A model that scored beautifully on clean sample images can fall apart on an actual workshop floor, out in a paddock, over a packing line, or in a yard at the wrong time of day. Lighting, angle, distance, movement, dust, glare, and whatever clutter is in the background all conspire against it, and none of that shows up in a demo shot in good conditions.

So before you commit real money, go and test the capture conditions on site. Put a camera where it would actually live and see what it gets across a full day, including the harsh afternoon light and the dusty afternoon and the shift where someone parks a ute in the background. More often than anyone expects, the fix that saves the project isn’t a better algorithm, it’s a fixed camera position and a decent light, so the system sees the same clear scene every time instead of guessing at a different mess each frame. Solve the seeing problem first, because no model rescues a camera that can’t make out the thing.

Take a packing shed on the Darling Downs grading produce over a belt. The trial footage from a hand-held phone in the morning looks great and the model finds blemished fruit nine times out of ten. Then it goes live and accuracy falls off a cliff every afternoon, and the culprit turns out to be the western roller door: for two hours a day the belt is half in glare and half in shadow, and the model is guessing. Nobody needed a smarter model. They needed a $400 LED array over the belt and a shroud, which is the sort of fix that never makes it into a vendor’s pitch because it isn’t clever enough to sell.

Detection is the cheap half of the project

Spotting the missing label, the damaged part, the open gate, the low bin, the defect, the worker without a hard hat, is the easy half. The value is entirely in what happens next. Does it ping a supervisor, raise a task, stop the line, attach the photo as evidence, drop a record on a dashboard? A detection that ends as a coloured box on a screen has just moved the work to whoever’s watching the screen.

That trigger, the wiring from “the system saw it” to “the right thing happened,” is where the project becomes worth having, and it’s mostly ordinary software and integration work rather than AI. The model spots it. The system files it against the right asset, job, batch or site, and sends it where it needs to go. Skip that and you’ve built an expensive way to generate loose photos nobody chases up.

Do the arithmetic before the pilot

Vision systems have a cost shape worth understanding before anyone gets excited. The hardware is cheap: an industrial camera, a mount, a light and a small computer might run a few thousand dollars per station. The sensor and vision software, the training, and the wiring into your other systems is where the spend lives, and there’s an ongoing cost to keep it accurate as conditions drift. So the question isn’t whether the technology works. It’s whether the miss you’re catching costs enough, often enough.

Work it from the miss. A load-check camera at the yard gate that catches an unstrapped item or the wrong pallet before the truck leaves is worth a lot if a miss means product returned from three hundred kilometres away, a redelivery, and a customer credit, call it $2,000 a time, twice a month. That’s roughly $50,000 a year of misses, and a system costing a fraction of that to build and run is an easy yes. But if the visual check you want to automate fails twice a year and costs $200 when it does, a camera system is a hobby, not an investment. Plenty of vision pitches skip this sum because the demo is impressive. Do it anyway, and make the vendor do it with you.

While you’re at it, budget for the boring failure modes, because they kill more deployed systems than model accuracy does. Lenses get dusty and someone has to own cleaning them. Cameras get bumped by a forklift and nobody notices the angle changed for a week. Seasons move the sun, so the scene that was fine in May is backlit in November. Night shift runs under different lights entirely. None of this is hard to handle if it’s in the plan; all of it is fatal if the plan assumes the camera looks after itself.

Let people rule on the hard cases early

In the early days, route the uncertain and the high-stakes detections to a person before anything acts on them automatically. That does three things at once. It stops a bad call from disrupting operations. It turns every human decision into a labelled example that sharpens the system. And it lets your staff watch how the thing reasons, which builds trust faster than any pitch, because they’ve seen where it’s right and where it isn’t.

This matters most exactly where the costs are lopsided: a false positive that bins good product or halts a line for nothing, or a false negative where someone could get hurt. Those are the cases you want a human on longest, and the ones where “the AI was pretty sure” is not an acceptable answer.

There’s a data reality behind this too. The events you care about most are usually the rarest ones. A line might produce one defective item in five hundred, which means weeks of footage contain only a handful of examples of the thing you’re training the system to catch. So expect a collection period before the clever part works well: cameras running, people reviewing, examples accumulating. The review loop above isn’t just a safety net, it’s how the training data gets built, and any applied AI plan that promises high accuracy from day one on a rare event is promising something the data can’t support yet.

Start with a scene you control

The strongest first projects have steady, controllable conditions: a fixed camera over a line, a proper inspection station, a known set of products, the same site shot the same way each time. Wildly variable outdoor scenes can absolutely work, they just need more data and a lot more testing before you lean on them, so they’re a second project, not a first one.

The regional use cases that fit are the ones that pair repetitive visual work with a clear next action. Checking PPE is on before entry. Counting livestock or stock items. Reading a gauge. Catching damaged packaging. Inspecting machinery for a known fault. Sorting crop or weed images. Confirming the right gear is loaded before a truck pulls out. Pick the first one because the deployment is achievable and the payoff is measurable, not because it makes an impressive demo. If you’ve got a repetitive visual check that keeps costing you when it’s missed, tell us the scene and the conditions and we’ll tell you honestly whether the camera can do it before anyone talks about models.