layman/terms
AI words

Computer vision

kum · pyoo · tur vizh · un

In shortComputer vision is AI that looks at photos or video and works out what's in them: faces, words, objects or what's happening.

A camera can see everything and understand nothing. Computer vision is what lets it say "that's a cat".

Show a toddler a picture book and they’ll point and say “dog!”, “ball!”, “car!”. Nobody gave them a list of rules for what a dog looks like. They just saw a lot of dogs.

Computer vision is AI that does the same with photos and video. It looks at the picture and says what’s in it: a face, a word, a car or a person falling over.

How it learns to see

To a computer a photo is just millions of coloured dots. Computer vision is trained with machine learning on millions of labelled pictures: “this is a cat”, “this is a stop sign”. Bit by bit it learns which patterns of dots usually mean which thing.

What it can do

Where you meet it

Unlocking your phone with your face. Searching your photos for “beach”. Self-checkout tills that spot what’s on the scale. Cars that brake when someone steps out. Doctors’ tools that point out a worrying spot on a scan.

Where the comparison breaks down

A toddler knows a dog drawn in crayon is still a dog. Computer vision can be fooled by an odd angle, bad lighting or a sticker in the wrong place. It only knows the kinds of pictures it was trained on.

How the words fit together

Computer vision runs on a neural network trained with machine learning. An AI that handles pictures and text together is multimodal. And the same skills that read pictures can help make fake ones: deepfakes.

Good to know

Your phone already does this every day: it finds faces to focus on, searches your photos for "dog" and copies text out of a picture.

When you'll hear it

“It uses computer vision to read the meter.” A camera takes a photo and software reads the numbers off it.

“The model detects objects.” It draws a box round each thing it recognises and names it.

“Face ID failed.” The phone's camera didn't recognise your face this time.

Say it in a meeting

“What could a camera spot for us that someone checks by eye today?”