With the current state of computer vision and AI, I would agree with GM. There's a lot of fast and loose talk about what AI can and can't do, but the reality is we're not even close to having general intelligence which would, at least in a rudimentary form, be necessary to safely navigate a car using only optical recognition.
> to say you can be a full Level 5 with just cameras and radars, is not physically possible.
Of course it’s possible. Humans manage to drive just fine without even radar. Optical recognition is clearly sufficient.
With that said, we clearly have not solved self-driving cars and it’s very possible that better sensors will improve the tech faster than smarter computing.
WOuldn't that logic suggest that self-driving tech dependent only on image recognition has a hard-limit at around the human driving expertise-level? If we truly want super-human driving skills (as everyone seems to expect from future cars), wouldn't that mandate a richer and better quality of data?
Computers have advantages in attention span, (lack of) fatigue, and reaction time. Potentially their performance could well surpass humans even without superior sensor abilities.
Not really. Dogs have roughly the same visual capacity as humans but are really shitty drivers nonetheless. Visual acuity doesn't necessarily correlate tightly with driving ability. It's possible that human-level driving is as good as it can get with just optical vision, but there's no evidence to support that. There may be vast untapped potential from pure optical vision. Honestly, if we just got self-driving cars as good as non-distracted, expert human drivers, they'd already look super-human compared to the average human.
Still, it's very likely that we can improve faster with better sensors.
> Dogs have roughly the same visual capacity as humans
While I more or less agree with the rest of your post, source needed on the quoted part. From everything I’ve read on the topic, there is significant differences in dog perception over humans (FoV, depth perception, motion detection, flicker fusion, etc).
If you limit image recognition to fixed binaural vision at a PD ~70; sure.
But cars can have more cameras, better infrared sensing, multiple exposure sensing, better field of view, quicker/better rotating ability and be couple more tightly with car speed to make more accurate predictions, adjustable known zoom.
So, if the human driving level were Deadshot + Jason Bourne + the Exorcist + Riddick + American Sniper + the Mentalist + Ghandi.
Umm... not without losing bet number one and the associated brand and image impact of selling thousands of cars that were sold as self-driving capable.
This is like Lockheed/Boeing execs saying that landing a rocket booster and reusing it was impossible. There's a clip of that somewhere which I can't seem to find right now.
19 comments
[ 3.6 ms ] story [ 50.3 ms ] threadOf course it’s possible. Humans manage to drive just fine without even radar. Optical recognition is clearly sufficient.
With that said, we clearly have not solved self-driving cars and it’s very possible that better sensors will improve the tech faster than smarter computing.
Still, it's very likely that we can improve faster with better sensors.
While I more or less agree with the rest of your post, source needed on the quoted part. From everything I’ve read on the topic, there is significant differences in dog perception over humans (FoV, depth perception, motion detection, flicker fusion, etc).
Actually, I’d argue that it is, but maybe my definition of “vision” is overly pedantic.
Vision is to seeing as listening is to hearing for me.
If you limit image recognition to fixed binaural vision at a PD ~70; sure.
But cars can have more cameras, better infrared sensing, multiple exposure sensing, better field of view, quicker/better rotating ability and be couple more tightly with car speed to make more accurate predictions, adjustable known zoom.
So, if the human driving level were Deadshot + Jason Bourne + the Exorcist + Riddick + American Sniper + the Mentalist + Ghandi.
Level 5 impossible without lidar: Tesla licenses the tech and buys the hardware like everyone else
Looks like two upsides on this bet.
I'm still skeptical that it can be implemented successfully, though, given that we still don't have Level 5 with lidar.