There’s little quality training data available to teach artificial brains, especially to perform image analysis. For supervised machine learning, you need lots of annotated data, annotated meaning a description of what’s in the picture and where exactly it is.
Today, such data are prepared by human annotators that draw bounding boxes or segmentation masks and write labels for every training image. Some things, like tractors, don’t even fit into a bounding box very well. Ideally, you’ll want a dynamic aspect as well, not just what it is, but what it’s doing, where it’s going, how it’s behaving.
There’s hardly any annotated data for different sensing modalities, such as infrared and LiDAR. Depth maps and point clouds require annotation in 3D space, which is a tough job for a human annotator.
Through our HySAIT (HyperSpectral Artificial Intelligence Training ) technology, we can help anyone developing artificial neural networks, by providing training image materials for deep learning. The data we can generate are beyond compare and exceed human annotation capabilities, both in their quality and quantity.
We are developing a hardware solution that can take the pain out of the image annotation process. Our system will be able to capture real-life scenery and generate annotation automatically without any human intervention - and at video rates or even in real time.
Our technology can automatically establish the so called “ground truth” about the objects or subjects within a 3D field of view – like their identity, distance, and orientation. And then apply that “ground truth” to the outputs of various sensors and imagers – visible, night vision, thermal, and even LiDAR! Make no mistake - these are real-life images, not simulations or synthetic computer graphics.
So we can supply terabytes of annotated training images, or even as annotated videos, for multiple imaging and sensing modalities simultaneously. So basically we’ve already done the hard work behind sensor fusion, because our data represent the same world, the same “truth” as observed through the “eyes” of different sensors.
Just the catch is that we need to capture this kind of data ourselves using our own hardware – we can’t apply such processing retroactively to any existing data.
Ultimately, we want to do crowdsourcing and aggregate training data in real time using a distributed network of data collectors. This will allow us to sell such data by monthly subscription to OEMs, Tier 1 suppliers or even large fleet operators.
We'll publish ads like this one, where we would compile a dataset that is equivalent to a year’s worth of driving experience in a particular country. You can then actually buy updates to the driving proficiency of your robocar and claim a discount from your motor insurance, because your car computer will now be a safer driver. So we can get insurance companies involved to speed up the adoption of our data.
Some other business models are possible like franchising of data collection operations in different countries.