4.5.23-AI-Ubicept

SEBASTIAN BAUER: Thank you very much, Ariana. My name is Sebastian. I'm the co-founder and CEO of Ubicept. Three out of our five co-founders have an MIT connection, so a professor, from a postdoc, and a PhD graduate.

What Ubicept does is really super-perception in very challenging environments. So all of us have cameras in our pockets, on our phones. And they operate reasonably well. When you have conventional lighting conditions, daylight or bright illumination, everything is fine.

But when you have a camera system on a moving platform that needs to operate in all kinds of environments, especially when there is fast motion, this is challenging for conventional cameras. Seeing in low light is very challenging for conventional cameras, and also brightness variations, super-dark, super-bright, changing over time or even in the same image simultaneously. And what we build is the computer vision platform on new types of image sensors that operate slightly differently from existing sensors. And that allows us to deliver much better results for computer vision.

This is an example of what I'm talking about. So what we see on the left is a driver monitoring system that also needs to be able to operate in low light when the car drives at night. And this is about driver drowsiness detection, detecting distractions and similar things. And obviously it's so dark that the conventional cameras, looking at the person's face, it's not able to pick up the view direction and the outline.

And our example in the sensor as are in the center, this is when we came out of stealth at CES three months ago, is the conventional automotive camera facing to the side of a car. And it's about detecting obstacles, different objects that are in the scene. And what we see here is an object detector running on the output stream of the conventional automotive camera.

And apparently there's a lot of false detections, false positives of cars in the scenes. There is no cars here. Another example on the right is QR code detection, where it's about detecting QR codes for logistics applications. And with this low light machine vision sensor, we're not able to resolve that QR code.

When I'm switching over to the next slide, we see really the performance difference. So what this is, it's a single photon-sensitive camera with Ubicept's computer vision and image processing algorithms running on top. So there you see the face of the person much more clearly.

In the example in front of this building, you see that there's a lot of pedestrians. There's zero cars visible there, and the object detector is reliably detecting these pedestrians. And on the right, the QR code detector is also able to see that QR code and decode it.

I also want to give some more insight into what we call super-perception. This is perception that is better as a human. And what you see here is the line that we call reliable eye threshold. So when you have a moving platform in an uncontrollable environment, it's about fast motion that needs to be resolved.

The other example is low light. And a third example are brightness variations. So just to add some numbers here, in order to reliably in all environments, you want to make sure that your system operates reliably above that threshold. So you want to detect very fast motion reliably. You want to be able to work in low light, and also you want to be able to have a huge dynamic range in order to be not distracted by brightness variations.

And what you see here for comparison are conventional sensors, CMOS sensors, the left-most bar, event cameras which are really good at resolving motion but are not that good in resolving low light. And human vision also is not really reliably working. Obviously all of us know when it's super-dark we don't see a lot.

And also the other thing is we are easily distracted, we as humans, by brightness variations. We can adapt to them inside a dark cave and outside in bright sunshine. But it takes us a certain amount of time to adapt to that.

Let's get more concrete here with that video here. This is an example of license plate recognition. There is many more applications, obviously. What we have here is the difficult scenario. It's low light and also there's pretty fast motion.

And on the left, you see a low-light camera. And on the right you see really very crisply resolved the car itself, even while it's moving and the license plate, for example, can be detected very reliably. Also the make and model of the car can be seen. This is something that a conventional sensor certainly can't do.

So what we are doing right now is we have an evaluation kit available. The camera sensor you see in the bottom center, that is from our hardware partner, and the big black box on the right is the compute box. And this combination gives you specs that really cannot be achieved by any other camera in the world.

We can do 500 frames per second output with 140 dB dynamic range in each frame, which is remarkable because conventional sensors, when you want to increase the dynamic range, the frame rate goes down. That doesn't happen with this evaluation kit. We can go down to 100 millilux brightness levels and also are looking forward to going even lower in the future.

Right now we are looking for partners in many different industries, kind of in this bucket, moving platforms and uncontrollable environments, but also industrial inspection, surveillance, defense, robotics, and many, many more. So if you want to learn more, please stop by our table. We have the prototype camera there and also can explain you much more in detail how we are exactly doing things. And looking forward to engaging with you. Thank you.