Ammar Saleem
Computer vision is at its best when a task is visual, repetitive, and currently limited by human attention. It struggles when the visual signal itself is weak or ambiguous. Knowing which side of that line a project sits on is the most useful thing to establish before spending anything.
Five settings where vision reliably wins
These tasks share a trait: the right answer is visible and well defined, even if the volume makes humans unreliable. A model does not get tired at the end of a shift.
| Setting | What the system does | Why it works |
|---|---|---|
| Quality inspection | Flags scratches, dents, misprints, and missing parts on a production line | Defects are visible and the volume overwhelms manual checks |
| Presence and absence checks | Confirms a component, label, or seal is where it should be | A simple, well-defined visual question asked thousands of times |
| Reading text and codes | Captures serial numbers, labels, meters, and documents (OCR) | Printed text is a strong, consistent signal |
| Counting and measuring | Counts items, checks dimensions, tracks fill levels | Objective answers that are tedious for people |
| Safety monitoring | Detects missing protective equipment or people in restricted zones | Clear visual rules and a high cost of missing an event |
Three settings where it struggles
Subtle judgement calls, scenes with poor or wildly varying lighting, and rare events with almost no training examples are where vision projects stall. None are impossible, but all need honesty about cost and accuracy.
- Subtle judgement: if a skilled human disagrees with another skilled human, do not expect a model to settle it cleanly.
- Uncontrolled conditions: outdoor scenes, glare, shadows, and changing light all degrade accuracy unless the setup controls for them.
- Rare events: a defect seen twice a year gives a model almost nothing to learn from, and needs a deliberate data collection plan.
Make it real: the physical setup
The hard part is rarely the model. It is lighting, mounting, edge cases, and the pipeline that turns a prediction into an action your systems trust. Design for the messy real world and the accuracy follows.
Consistent lighting, a fixed camera position, and the right lens and resolution often improve results more than a better model. Where latency, privacy, or connectivity matter, run the model on an edge device next to the camera, and keep a central place to monitor and update it.
How to estimate the return
The return comes from three places: labour no longer spent on manual inspection or data entry, defects caught earlier (before they become rework, returns, or recalls), and consistency that sampling-based checks cannot provide.
Weigh that against cameras and edge hardware, data labelling, the build itself, and ongoing monitoring and retraining. A short feasibility study on a few hundred of your own images is the cheapest way to replace guesses with numbers.
Getting started
- Pick one inspection or capture task with high volume and a clear definition of right and wrong.
- Collect and label a few hundred representative images, including the awkward cases.
- Measure model accuracy on those images before committing to a full build.
- Fix the physical setup: lighting, camera position, and triggering.
- Run in shadow mode alongside the current process and compare results.
- Go live with human review for low-confidence cases, then monitor for drift.
Use vision for visual, repetitive, attention-limited work. Be cautious where the signal is weak, control the physical setup, and prove accuracy on your own images before scaling.
Frequently asked questions
Common uses include automated quality inspection, checking that parts or labels are present, reading text and codes with OCR, counting and measuring items, and monitoring safety compliance such as protective equipment.
For well-defined visual checks at high volume, it is usually more consistent, because it does not tire or vary between shifts. For subtle judgement calls, skilled people often remain better, and the best systems route uncertain cases to them.
Often less than expected. Pre-trained models can be adapted with a few hundred well-labelled images for many tasks. Rare defects are the exception and need a deliberate plan to collect examples.
Working on something like this?
See how our Computer Vision practice works, or talk to us directly. No pitch.
