Human eyes don't work anything like cameras. The eye processes information that isn't just intensities of photons and their wave lengths—which is the only information collected by cameras. See Walter Pitt's "What the Frog Eyes Tell the Frog Brain" for an introduction: 'Fundamentally it shows that the eye speaks to the brain in a language already highly organized and interpreted, instead of transmitting some more or less accurate copy of the distribution of light on the receptors.'

A simple abstraction here: we put two high-quality monitors before our eyes and get all the data from them. Can we continue driving this way? Absolutely. This means that no matter how our eyes actually work, an array of RGB pixels is enough. I'd go further and say that a 64 x 64 grayscale camera at 10 fps is enough to drive a car. The real magic happens in the brain that reconstructs the 3D model and predicts what happens next.

> "64 x 64 grayscale camera at 10 fps is enough to drive a car."

If you had vision that bad, the State of California would not let you get a drivers license, they require at least one of your eyes to be 20/40. Anyway, the point of the GP is that the brain isn't the only piece of wetware doing image processing, but more to the point the current state of the art CV really is not up to the task.

And it really doesn't seem like it's going to be there anytime soon.

That is an analogy not an abstraction but if you had read the paper or the single quote I provided, you would know that even relatively simple eyes (such as those of the frogs which cannot see still objects) don't only process a distribution of light intensity and frequency. The eye measures other information that is lost by the camera including photon phase, arrival time, polarization, orbital angular momentum, linear momentum, and probably many more measurements. More so, this is done by specialized organelles of the eye rather than by the brain.

It's great that eyes can detect polarization and orbital momentum. My point is that none of that info is needed to drive a car.

That's true, but the main point is that current software still relies on sampling. Eyes don't sample.

And this is irrelevant also, because we can introduce sampling the way I described and can continue driving just fine.

I've thought about it and yes this does make sense.

Humans can do a lot of things we don't understand how to convince software (...or other humans who don't intrinsically learn it very well) to do.

Lidar lets you get a high degree of accuracy for certain variables in your model - because you can know fairly accurately when and how strong the signal you sent out is, you can have a fairly high degree of confidence when you get reflection of how far away the object is, and feed that into your persistent idea of the world, versus having to use a bunch of heuristics to try and extract that from purely audiovisual data.

(That's my understanding, at least - I don't work in that industry.)

A good way IMO to think about it is by looking at airplanes. Birds fly by flapping their wings to generate thrust, but planes use turbines to generate thrust.

Just because humans use vision to drive, doesn't mean that we should restrict ourselves to vision for cars. Also, as another comment points out, our visual system does a lot more than just 'see' in the literal light/color processing sense that a camera does

Because current AI does not seem as good at generalizing as humans?

Since level 5 driving has not yet been achieved, aren't we talking about future AI, not current AI?

Tesla's marketing team wants people buying cars to believe their Tesla will support Level 5 automation sometime within the usable lifespan of that car. They're not bold enough to claim it's a near future inevitability, but they're leaving that implied possibility open to get consumers excited about it.

Google/Waymo has obviously been throwing a ton of resources at nailing computer vision, their recaptcha that's plainly training such a system is doubtlessly just the publicly visible tip of an iceberg. Yet they're still investing into LIDAR as well. If Google/Waymo is bearish on a pure CV solution, I don't see any justification for Tesla thinking they can do it. Not in any near future.

https://arstechnica.com/cars/2017/01/googles-waymo-invests-i...