Computer Vision in New Zealand

The camera and the lighting decide the project

This is the thing most briefs miss, and it is worth more than any choice of model.

A fixed camera, controlled lighting and a consistent background turn a hard problem into an easy one. Variable daylight, a camera someone bumps, steam, dust and a background that changes turn an easy problem into an impossible one.

Practical implications:

• Controlled indoor light beats natural light every time. Shadows and glare are what break models.
• A fixed mount beats a handheld device, because everything the model learned assumed a viewpoint.
• A plain, consistent background is worth more than a better model.
• Position matters more than resolution. A cheap camera in the right place beats an expensive one in the wrong one.

Ask each supplier what they would change about your setup and whether that is in the quote. Anyone who wants to talk about models before cameras is skipping the part that determines whether this works — and lighting is usually a few thousand dollars against a project of a hundred.

Prove it on your own site

Accuracy figures from a supplier's samples tell you about their samples.

A proof of concept should use images captured in your real conditions, including the bad ones: the wet morning, the end of a shift when the lens is dusty, the batch that arrives packed differently, the light at 5pm in June.

What you should get back:

• Accuracy per category, not one headline number.
• The cases it gets wrong, shown as images so you can see the pattern.
• An honest statement of what would need to change on site to improve it.
• A recommendation, including "not with these conditions".

Expect $20,000 to $45,000. It is the cheapest way to find out whether the business case is real.

Labelling is your cost, and it is larger than you think

The model learns from examples that somebody has marked up, and only your people know what a Class 2 defect actually looks like.

Expect to supply several hundred to a few thousand labelled images depending on the task, and expect the labelling to take real time from someone experienced.

Two things make it go badly:

• Disagreement. Two experienced graders will label the same item differently. If your own people cannot agree, the model cannot learn a consistent rule — and discovering this is itself valuable, because it means your manual process is inconsistent too.
• Rare cases. The defect that appears twice a year is the one that matters and the one you have no images of. Start collecting now, before the project.

Budget the labelling explicitly. It is the most commonly underestimated line in these projects.

Decide where it runs

On-site processing is right when decisions must be immediate — rejecting an item on a moving line — when connectivity is unreliable, or when video volumes make uploading impractical. That covers a great deal of rural and industrial New Zealand.

Cloud processing is right when images are few, a second or two of latency is fine, and you would rather not maintain a computer in a shed.

Whichever it is, ask what happens when the connection drops. On a packhouse line the answer needs to be that nothing stops.

Plan for what happens when it is unsure

A system that always produces an answer will produce confident wrong ones. A system that flags uncertainty and routes it to a person is worth far more.

Settle during quoting:

• What is the confidence threshold, and who set it?
• What happens below it — hold, flag, or pass to a person?
• Who reviews, and how quickly, without stopping the line?
• Do their corrections feed back into the next training round?

The threshold is a business decision. In grading, a false accept and a false reject cost very different amounts, and the threshold should reflect that rather than being left at a default.

If it can see people, it has obligations

Cameras capturing people bring the Privacy Act into the project, whether or not the people are the point of the system.

Practically: tell people the cameras are there and why, collect only what you need for a stated purpose, decide how long footage is kept and delete it, control who can view it, and think very carefully before anything that identifies individuals rather than counting them.

Staff monitoring in particular is an employment matter as much as a technical one. Raise it early — it is much easier before installation than after.

Comparing five quotes

Read for:

• Who visited, or asked for images from your actual site?
• Who talked about cameras, mounting and lighting before models?
• Who was explicit about how much labelling you would have to do?
• Who included on-site processing and the connection-drop case?
• Who described what happens when the system is unsure?

Describe what you want counted, graded or read, where it happens, what the light and the environment are like, and what a mistake costs in each direction. Suppliers who have shipped this will ask for photographs of the site before they quote anything.