Computer vision
Spot it in the picture
A model trained on labelled images finds an object in a camera view and draws a box round it.
Need what you build to see, understand and create?
AI lets an app, headset or web system recognise what a camera sees, understand what people write and find patterns in data. In your training experience, it could recognise the kit a trainee points at.
The idea A virtual reality training experience for our teams
See it work
Four kinds of AI, shown not named
Each example is a drawing of what that kind of model does, not a live model.
“Book my team on the forklift refresher next Tuesday.”
- Intent
- Book a course
- Who
- My team
- Course
- Forklift refresher
- When
- Next Tuesday
Natural language processing
Understand the request
A language model works out what someone wants, and the details that matter, from typed or spoken words.
Prompt A safety sign for the loading bay
Generative AI
Make something new
A generative model learns from examples and makes new text, images, sound or data like them.
Machine learning
Find the pattern
Machine learning finds patterns without being given the rules. Left to itself, this model sorts the points into three groups.
What it is
AI inside the thing people already use
-
One of our developers at work in Unity.
AI is software that does jobs we used to need a person for: recognising a product, understanding a sentence, spotting a pattern in thousands of records. We build it with you into apps, headsets and web systems, inside the experience people already use.
It already does this work in healthcare, finance, manufacturing and entertainment. In your training experience, it might find the steps where most people hesitate.
Image recognition
Teaching a camera what it is looking at
Image recognition is the part of computer vision that identifies objects, people and actions in an image. A model learns it from labelled examples, often millions of them, so their quality and variety decide how well it works.
| Where | What it recognises |
|---|---|
| Augmented reality | The image, object or surface to place content on |
| Manufacturing | Faults found during quality inspections |
| Healthcare | Signs of disease in MRI scans, X-rays and CT scans |
| Retail | Products, to help keep track of stock |
| Self-driving vehicles | Pedestrians, road signs and other vehicles |
Synthetic datasets
Training images that arrive already labelled
A model usually learns from thousands of photographs, each with a box drawn round the object by hand. That is slow, and some subjects are awkward or sensitive to photograph.
With Unity’s Perception tools we build the object and its surroundings in 3D instead, and render as many images as the model needs, each labelled as it is made.
-
Unity Perception output: a render, the same render labelled, and its label data.
By hand
- Photograph the object in every setting
- Draw a box round it in each image
- Repeat for thousands of images
Generated in Unity
- Build the object and its surroundings once
- Set the lighting, clutter, quantity and object
- Render thousands of images, each labelled as it is made
For a healthcare client, we built a Unity app that makes labelled images of their at-home finger-prick blood test kits: thousands in minutes, where labelling by hand could take weeks.
Real photographs still matter. Unity’s published tests found a model trained on synthetic and real images together was more accurate than one trained on either alone. The same approach can train a model to spot products, actions or facial expressions.
How we work
From first idea to a model that works
-
Find where AI helps
We look at the idea with you and agree where AI adds something people will notice, and where a simpler build would do.
-
Gather or generate the data
Together we collect real examples and, where they are slow to gather, generate labelled ones in Unity.
-
Train and test the model
We train or adapt a model and test it with you against the cases that matter most.
-
Build it in and keep it working
We build it into your app, headset or web system and look after it with you after launch.
Immersive AI
Want a character people can talk to?
Conversational characters, AI presenters and training that responds to each person live in Immersive AI.
Explore Immersive AIQuestions
Good to know
-
What kinds of AI do you work with?
Computer vision and image recognition, natural language processing, machine learning and generative AI, often more than one in the same project.
-
What is synthetic data?
Images made in a 3D engine that look like photographs and come labelled automatically, so a model can be trained without labelling every image by hand.
-
We are not sure AI is the right choice. Can you help?
Yes. Tell us what you want people to be able to do and we will work through it with you, including whether a simpler build would serve the idea better.
Your idea
Talk it through with us
Tell us what you are trying to explain, launch or change. We will work out with you whether this is the right medium for it.
Also in Digital Solutions
Other ways to tell it
-
Gamification
Your training is built and you can see how people use it. Now you want them to finish it, and to go again for a better result. Gamification gives them a reason to.
-
Web systems
Most apps and experiences need a system behind them to change the content, manage who has access and keep the data safe. Edit something once in a web system we build with you and every device shows it.
-
Experience Analytics
Your virtual reality training is live. Now you need to know who finished it, how they scored and which part held people up.
Your turn
What’s your idea?
Tell us the idea you are trying to explain, launch or change, and together we will find the right way to tell it.