Tesla is going against the grain with its new Tesla Vision system that ditches the radar and Lidar used by other automakers to make self-driving possible. It recently stopped equipping its new cars with anything other than cameras, yet the company is adamant that through a combination of cameras and a neural network working in real time, it is just as good as with radar and Lidar.
We were able to previously experience how older versions of the Tesla Full Self-Driving (FSD) saw the world and how they understood it, but those no longer apply. Tesla Vision perceives the world (and depth in particular) quite differently, as exemplified in this video uploaded by a well known Tesla hacker who goes by @greentheonly on Twitter
We’re not really sure why the resolution is so low and whether this is a software or hardware limitation. Probably the latter given the fact that it is way below the main camera’s native resolution. Tesla will further improve and enhance its new camera-based system once it launches FSD Version 9, which should be in a few weeks, according to company CEO, Elon Musk.
And if you were wondering how the neural network part of the equation works and how it learns, check out this older except from a Ted talk by computer science researcher Andrej Karpathy, who is also the director of artificial intelligence and Autopilot Vision at Tesla. In this video he explains how a neural network works and if you want to have the complete picture, you may want to watch the entire talk, not just this specific part about depth perception; the resolution shown during the presentation is much higher than in the video uploaded by the hacker.