Aug 4, 2022
AI/ML Simulation ServicesLet Luxonis work for you
The OAK line of products is the heart of Luxonis. They are the building blocks for our customers to create and invent so many amazing things, and offer an unmatched degree of performance and accessibility. And when you’re working with the best products out there it sometimes unlocks the ability to tackle problems that you otherwise wouldn’t dream of. Which then creates a problem in and of itself: complex problems can take a really long time to solve. That’s where we come in.
With DepthAI, Luxonis Has Your BackWe work hard every day to develop resources our customers need to support themselves. If you’re not familiar with everything our DepthAI Documentation pages have to offer, be sure to take a look. DepthAI covers everything from setting up your very first OAK camera, to reviewing concepts like spatial AI and on-device programming, to API tutorials, to hardware specification and functionality overviews. There’s a lot there. We even link you directly to a wide range of demonstrations and models for you to install and use in just a few clicks. We’re also always there on Discord to answer any questions you may have. But if you need support that’s more “soup to nuts,” we’re there too, in particular when it comes to dataset development for artificial intelligence (AI) and machine learning (ML) models.
Synthetic SympathizingWe get it. Creating and labeling datasets for AI and ML applications is no small task. Such datasets often require thousands of images for even very narrow use cases, and even more than that for complex or multi-purpose use cases. These projects can be labor, time, and financially intensive–all at once. But we have your back. Through the utilization of synthetic components to supplement real-world images, it’s possible to generate simulations for training models that offer a step-change in speed and efficiency. With the old approach, it can take months of effort to capture, transmit, and store terabytes of data. That’s expensive. And what’s even more expensive? It can cost upwards of $1+ per image to properly label that data so it can be effectively used to train a ML model. With the new approach–with Luxonis–all this can be done at a fraction of the time and expense. The whole point is to start using datasets as quickly as possible to solve actual problems, and synthetic supplementation helps make that happen. AI/ML/simulation services from Luxonis can assist at any point along the way. We can help research and obtain synthetic dataset components if they don’t already exist. We can set up lightweight data models including labeling, training, and integration. We can model deployment and testing within specified environments, as well as offer model evaluation and performance to identify areas for improvement. If you’re looking for dataset support services like this, look no further. Email us at: [email protected] to ask about your specific project.
The Grass is Always GreenerA great representation for the kind of services we’re talking about can be seen in the work done with our customer, Greenzie. Greenzie is an autonomous lawn mowing company that came to us with a problem. They wanted to deploy their fleet of mowers to fields, parks, and lawns all over the country, but needed a way to ensure that mowers would only cut what they’re supposed to. Driving over objects like rocks or sprinkler systems could not only damage the mowers, but their customers’ property as well. What was the solution? It would be too vast of a ML project to attempt to train a model to identify every possible object a mower could impact, so our teams collaborated and found a much simpler solution. We realized that the core of the issue was for mowers to know when to be “blades on” and when to be “blades off”. All grass within the permitted zone should be semantically segmented as “blades on” while all other objects regardless of type should trigger “blades off”. With this approach in mind, we set about generating datasets to begin training their models. Scenes in Unity were developed based on real locations, and then modified with challenging synthetic elements like rocks or fences. These scenes were then used to supplement data collected from Greenzie’s real life projects to form a more robust and comprehensive grouping of scenarios to learn from, all of which came together vastly faster than attempting to use real-world images alone. Then, based on the success of correct “blades on” and “blades off” identifications, additional scenes and synthetic elements were generated to refine problem points and greatly improve confidence intervals. Here we see a set of example synthetic images:
A scene developed in Unity with hazards to the mower (rocks).
The scene with grass semantically segmented as "blades on."
The scene with hazards additionally semantically segmented as "blades off."
And now let's compare those synthetic images to real-world performance:
A challenging area with sidewalk, trees, and a manhole cover.
An even more challenging example with bright blue sky, background objects, and a barely visible sprinkler head.
Synthetic supplementation was key to Greenzie’s success and is allowing them to continuously improve their dataset quality. And by contracting Luxonis’ AI, ML, and 3D artist specialists to help with the process they were not only able to save time and resources, but also redirect their own efforts elsewhere, allowing for better project efficiency across the board. So remember, while dataset development is never easy, it doesn’t have to be a barrier to progress. Work with our experts, and get your data working how you need it and where you need it. Fast.