> relying on in-car cameras rather than the radar and sensors ... It is betting that its approach will be cheaper
Is this due to reduced processing requirements, reduced sensor costs, or what?
Is this due to reduced processing requirements, reduced sensor costs, or what?
But humans in a car have a massive advantage over little cameras that no one seems to discuss much: we have two sensors (eyeballs) mounted on a servo (our head) that can move around and is looking through a truly monstrous aperture (the windshield), and that aperture is equipped with fancy cleaning devices (wipers and cleaning fluid spray), and the car’s operator is motivated to clean the windshield and maintain the windshield, wipers, and spray system to be able to see.
A Tesla car has little tiny camera lenses that are every bit as exposed as the windshield but don’t have all the compensating machinery.
Go stick a pair of nice cameras on a three-axis servo mount with a range of motion of a whole foot (or a camera array and no servo), stick that two feet behind the windshield, train it well (use that massive parallax!) and I’d believe the result would be competitive in performance but definitely not cost. Also the car would lose an entire seat.
Or use radar and lidar and achieve super-human performance.
Fir what it’s worth, the military was and is fully aware that lidar and similar tech can outperform human eyeballs in “battlefield conditions”, and I’m aware of old DARPA projects to do things like pulsed laser range-gated imaging to see through fog and such. (You still get attenuation and scattering, but you can mostly disambiguate the additive signal from fog from the stuff behind it.) Lidar can do something similar. Humans can move their head to acquire more data. Little cameras are at the mercy of the fog and can only use fancy image processing to try to compensate.
At the very least they couldn't they detect imminent collisions with pedestrians? walls?
----
My non-Tesla has a radar for cruise control and imminant crash detection — while annoying at first, it is unfailable. As I've aged (¡perfect driving record!) my thought process has gone from more horsepower! to more safety!. Now when my car beeps at me I'm more-inclined to listen cautiously.
My active cruise control works even when set to 99mph (e.g. it will slow me down upon approaching another vehicle). I'm sure it'd work at higher speeds, but won't "set" above 99.
And there should be some criminal liability since people have died.
Why even bother when we can make artificial eyes that see depth? The price of LIDAR has plummeted and will continue to plummet. We already know that it works really, really well for self-driving with today's available compute+data.
Eh, I mean I think that that’s necessary, but maybe not sufficient. Eyes are _really good_.
Given the number of shortcomings of human vision why shouldn't our self driving cars be designed to have better than human vision, especially if the goal is to not get into crashes. Humans, with human vision and human object tracking skills, and human reaction times get into crashes all the time. Shouldn't we want better and more sensors, which would lead to few fewer crashes, simply because better sensors have better data available?
Cheaper than that and you're making significant trade-offs.
It's also a bit of a false analogy. Cameras don't really work like human vision. We do things like mesopic vision that simply aren't possible with current sensors. We have massively high resolution. We have async "pixels" that can respond immediately without waiting for a framing signal. Our brains process color in truly weird ways.
It's not like there's some physical law preventing computer vision from being better than human vision, but it's an incredibly difficult engineering problem that we've spent the better part of a century attacking without clearly winning.
If you have reliable depth information, ala LIDAR, you'll be able to know that there's nothing actually there.
Lidar will likely be outlawed anyway as it burns your retina. Dare you to put your eye (or cellphone camera) next to a waymo sensor for 10 sec and see what happens.
HW3 has three front-facing cameras that are not only too close together to provide binocular vision with adequate disparity, but also have different fixed focal lengths making them unable to establish binocular focus even if they were far enough apart [1]. HW4 has two front-facing cameras with the same limitations.
This is of course ignoring the fact that humans with visual acuity comparable to the HW3 cameras would almost be legally blind and not meet minimum vision requirements to operate a motor vehicle in, I believe, every state. HW4 cameras are better and you would only be unable to meet minimum vision requirements to operate a motor vehicle in most states, including California and Texas.
[1] https://en.wikipedia.org/wiki/Tesla_Autopilot_hardware#Tesla...
https://www.laserfocusworld.com/blogs/article/14040682/safet...
Many of the class 1 lidars do damage camera sensors, is your eye really that much more resilient?
The short exposure is putting a lot of faith in these device manufacturers. How do they guarantee short duration, are you confident that each manufacturer has sufficient electronics to stop the laser if the spinning slows slightly due to dirt, low power or a crash? Should the exposure limits be the same during day vs. night when your pupils are dilated?
So, safe for human eyes but deadly for a camera.
Most of the problems I have now are things like lane selection or turning into the wrong parking lot, which seem solvable in software given enough time, and the Robotaxi project should ramp up the urgency on that front