FAQ

Your QUESTIONS

How Do I Get Started With Machine Dreams?

Give us a call! Every machine learning project is unique, but we can help design a simple Proof of Concept project with you for a fixed cost.

The PoC gives us a tightly focussed problem to work on, and we’ll work with you to pick a need that is immediately useful to your business. At the end of the PoC we are able to show results and levels of accuracy for detection of whatever it is you’d like to trial, so you can make informed decisions about going forward with synthetic data for your machine learning projects.

If you don’t know what you need, or don’t know whether you have the right data – don’t worry. We can help with everything from data acquisition and image capture through to eventual deployment and software development.

What Is Object Recognition?

The term “Object Recognition” comes from a discipline called Computer Vision – software, training data and algorithms that enable a computer to “see” and recognise an item, artefact or event in photos and video.

When trained, an object recognition model is able to examine photos and video content and report whether the things it’s been trained to recognise are present in the content, classify them and report their location in the image or video.

Computer Vision is itself a subsection of field of machine learning.

What can Machine Learning or Object Recognition do for me?

This technology can be used for many things including:

  • Tracking Assets or items
  • Finding things in large or complicated spaces
  • Predictive Maintenance – detecting defects or signs of breakage early, so you can fix things before they cause catastrophic damage or shutdown of a pipeline
  • Automated and remote inspection – get all your assets inspected effectively on a constant ongoing basis without sending teams out to review them, and ensure that when you send a repair crew it has all the knowledge and equipment it needs
  • Audit trails – scientifically proven accuracy in reporting along with geocoded and timestamped data to prove an inspection was properly performed, helping with liability and reporting issues

These results can be achieved with SaaS platform subscriptions, on-premise or on-vehicle software deployments, and even at the edge on Internet-of-Things devices and smart phones.

There are many more uses of this tech, and the field is just getting started, so get in touch and we can help you determine whether machine learning is right for you, and where it can save you time and money.

How do I get started with Machine Learning?

Machine Learning is a broad topic, covering many different techniques, approaches and technical architectures, so it can be a complex field to approach to begin with.

If you’re looking to start a machine learning project, learn more about how machine learning can be used or simply learn about the discipline itself – get in touch. We can provide presentations tailored for every background, whether you want a high level overview or a deep dive into the tech.

In addition there are several online courses, ebooks and tutorials available that we can help you and your team navigate to find what you need.

How Does It Work?

Generally a machine learning project for object recognition (detecting things in images and video) has four parts:

  • 1) Image Recognition

You gather lots of photos of the things you want the computer to detect, from as many different angles as possible, in different environments, backgrounds and lighting conditions.

  • 2) Labelling

You draw boxes around the part of the image you want the computer to look for and categorize them.

  • 3) Training

You turn all this data into formats the machine learning training software can understand, and set the machine to processing it. (This is the “learning” stage).

  • 4) Detection

You put the resulting “machine learning model” into some software so that it can look at images or videos and report back whether it detects what it’s been trained to look for.

When these stages are all completed successfully, you can just feed images from your reporting / on-site cameras, drones, helicopters etc into the software and it will tell you what it detects, with a level of “confidence” (IE “I am 98% sure there is a broken traffic light at position X,Y in this photo”)

What If I Don't Have Enough Photos?

That’s where Machine Dreams’ simulation engine, Morpheus comes in.

We use video game engine technology to generate “Digital Twins” of the items, assets or defects you want to detect, making them photo-realistic and creating systems that randomize things like dirt, wear and tear, rust, cracks and more.

Morpheus can then mix and match, swapping “clean” artefacts for defective versions, randomizing the effects of cracking, peeling paint, corrosion etc and switching configurations, environments, lighting, camera angles and more.

From these combinations, Morpheus generate super high-resolution synthetic “photos” of your assets and defects in as many variations as required. At the same time, Morpheus generates all the training data and pixel perfect labelling and segmentation information in order to effectively train a highly accurate machine learning model.

So no photos, no bother!

How does the Machine Dreams approach save me time and money?

Because we use synthetic imagery and generated data, our process skips the costly and time consuming Image Capture and Labelling steps entirely, providing more accurate training data and infinite variations of photos needed to make a machine learning model more accurate.

Automation of the machine learning process itself also saves time and money, and means you don’t need to engage or hire specific machine learning expertise yourself.

In short – we can skip phases 1-3, replacing them with our Digital Twinning and automated processes, saving you 70% or more of the time and cost of the traditional way of developing machine learning models.

What kinds of things can we detect?

If it can be seen, it can be simulated – and when we simulate it, we can teach the machine how to detect it in real photos and video.

This includes different forms of spectroscopy – infra-red/thermal, ultraviolet, LIDAR, point clouds and even 4D data like ultrasound video.

Morpheus can also work in the abstract and on two dimensional content like medical scan imagery and DNA PCRs, and even at the micro-level for things like cellular photography, microchips and other things not visible to the human eye. If a camera, microscope, FLIR or other specialised device can see or visualize it, Morpheus can simulate it.

What Is Morpheus?

Morpheus is the name of our simulation engine, built using video game engine technology. Using a standard we have developed, it allows us to throw digital twins of assets and artefacts, environments (scenes), and defects together, randomizes everything and produces high resolution “synthetic” photographs that can be used in place of the millions of real photos needed to create an accurate machine learning model.

We also create special pieces of game development tech to simulate and randomize effects like cracking, rust, peeling paint, graffiti, corrosion, warping and more, to ensure millions of variants of defects can be generated for detection training.

If you’d like to know more about the technology or processes involved in Morpheus’ use and creation, please get in touch – we’d love to show you and your team a presentation on how you can use Morpheus to achieve your machine learning goals.

Do you have more questions?