Biologically-Inspired Capabilities 

It’s often said that imitation is the highest form of flattery. When it comes to solving problems, humans have long looked to nature for inspiration. From seemingly simple concepts like Velcro(R) being inspired by the way burrs stick to dog fur, or feats of engineering like the bullet trains whose noses were fashioned after the beak of a Kingfisher to reduce the effects of shock waves as the train passes through a tunnel.

In these - and many more - cases, nature gives us a head start that is so effective it can’t be overlooked. And so it is with our OAK devices running Stereo Neural Inference to determine object distance, very similar to how our eyes do.

How We Do It

When we look at something, each of our eyes focuses independently on the object and the angles made as our eyes servo to the object allow us to triangulate the distance to the object. We accomplish the same task on an OAK device by running multiple neural networks in parallel on two or more cameras, but instead of algorithmically determining the depth of each pixel in the scene (which OAK devices can also do simultaneously), we use image AI to identify specific features of an object from the perspective of each of the camera’s ‘eyes’ and then calculate distance by measuring the pixel disparity between those features.

Note the relationship between pixel disparity (X axis) and distance (Z axis) in the following example. When the subject is closer to the camera, pixel disparity increases, when it is farther away, it decreases.

Why This Matters & How To Use Stereo Neural Inference on Your OAK

The power of AI is that human-generated algorithms have limits; humans simply can't write enough algorithms fast enough to cover all possibilities. AI overcomes these limitations by running trillions of experiments a second along with image recognition and object detection functions to learn what the correct algorithm is. With Stereo Neural Inference, OAK devices bring the power of AI to depth estimation.

Things that would hamper traditional depth estimation - and compromise your entire project - can now easily/readily be solved with Stereo Neural Inference. Examples of things that would compromise traditional stereo depth functionality, yet are easy to solve with stereo neural inference include (but are definitely not limited to):

• Finding the 3D location of shiny/reflective objects
• Finding the 3D location of objects with subtle features/textures.
• Finding the 3D location of very small objects

Check out the video below to see just how easy it is to get started with Stereo Neural Inference on an OAK-D-Lite device. It’s easier than you think!