Computer vision projects fail on physical problems far more often than on technical ones. The model is rarely the hard part. Lighting, camera placement, and whether anyone acts on the output decide almost every deployment.
Here are five uses that reliably pay for themselves, and what each actually requires.
1. Visual quality inspection on a production line
What it does: checks every unit for defects, missing components, wrong labels, incorrect fill levels or packaging faults.
Why it pays: human inspectors are good for about twenty minutes before attention degrades, and they cannot inspect every unit at line speed. A camera inspects all of them, consistently, at 3am.
What it needs: consistent lighting and a fixed camera position, which is genuinely the hardest part. It also needs several hundred examples of each defect type, which means someone must collect and label them, and defects are by definition rare. Budget for that collection period before expecting results.
Best fit: pharmaceutical packaging, food and beverage filling, textile fabric inspection, electronics assembly.
2. Safety compliance monitoring
What it does: detects whether people in a defined zone are wearing helmets, high visibility vests or safety boots, and whether anyone has entered a restricted area.
Why it pays: one serious incident costs more than the system, and for exporters, documented safety compliance is frequently a customer or certification requirement.
What it needs: cameras already installed in most cases, plus a decision about what happens on a detection. An alert to a supervisor works. A log nobody reads does not.
The important part: decide the policy before deployment. This system watches workers, and introducing it quietly creates resistance that outlasts any benefit. Framed as protection and applied consistently, including to management, it is accepted.
3. Counting and measuring
What it does: counts items on a conveyor, bags on a pallet, vehicles entering a yard, or people entering a store.
Why it pays: counting is tedious, error prone and constant. Automated counts also produce a record that can be reconciled against dispatch documents, which is where discrepancies get found.
What it needs: a controlled view of the counting point. This is one of the most reliable computer vision applications because the task is narrow and the accuracy is easy to verify.
4. Number plate recognition
What it does: reads vehicle plates at a gate, logs entry and exit, and matches against an expected list.
Why it pays: automates gate management at factories, warehouses, housing societies and parking facilities. Combined with weighbridge data it produces a complete record of what came in on which vehicle at what time.
What it needs: a camera at the right height and angle, and adequate lighting at night, which is where most installations fail. Pakistani plate formats vary considerably, so the system needs tuning on local plates rather than an imported configuration.
5. Retail shelf and display monitoring
What it does: checks whether products are on the shelf, whether the planogram is followed, and whether competitor products have taken your space.
Why it pays: for brands selling through retail, out of stock on the shelf is invisible from head office and directly costs sales. Field staff photographing shelves and a system doing the checking replaces a manual audit that happens quarterly at best.
What it needs: a disciplined photo capture process by field staff. The technology works; the compliance of the people taking the photographs is the variable.
What decides success, in order
- Lighting. Consistent, controlled lighting improves results more than any model change. Budget for it explicitly.
- Camera placement. Fixed, correct angle, correct distance. A camera that moves invalidates the training.
- Enough labelled examples, especially of the rare cases you actually care about.
- A defined action on detection. An alert to a person, a line stop, a log entry with an owner. Detection without action is a dashboard.
- A feedback loop. When the system is wrong, someone marks it, and that correction improves the next version.
Realistic costs
- Pilot on one line or one gate: 600,000 to 2,000,000 PKR including cameras, lighting and edge hardware.
- Cameras: 15,000 to 150,000 PKR each depending on resolution and whether industrial rated.
- Edge computing device to run inference on site: 80,000 to 400,000. Processing locally is usually preferable to sending video to the cloud, on both cost and bandwidth.
- Running cost: maintenance, retraining as products or conditions change, and cleaning the cameras, which sounds trivial and is a real operational task in a factory.
How to start without wasting money
Pick one narrow problem with a measurable cost. Not “quality inspection” but “missing cap detection on line three”. Run a pilot on one line for two months and measure against the current manual result.
Vision projects that begin with a broad ambition produce a proof of concept that impresses everyone and gets deployed nowhere. Projects that begin with one measurable defect on one line get extended.
For the wider question of where automation genuinely absorbs work, see our map of what AI actually takes over, and what running these systems costs monthly.
Frequently asked questions
Do we need expensive cameras?
Usually not. Lighting and placement matter more than resolution for most tasks. Industrial rated cameras matter in harsh environments.
How much training data is needed?
Several hundred labelled examples per defect type is a reasonable starting point. The constraint is usually collecting examples of rare defects, not the labelling itself.
Can it run without internet?
Yes, and it usually should. Running inference on an edge device on site avoids bandwidth cost and keeps working during outages.
What is the most common reason these projects fail?
Inconsistent lighting, cameras that get moved or dirty, and no defined action when something is detected. All three are operational rather than technical.
Ezitech builds computer vision and AI systems for manufacturing, logistics and retail, including edge deployments. See our AI solutions or describe what you want to inspect.