5 ms·

I agree in a big way here, I have no particular speciality with autonomous driving but I have seen enough media coverage of Musk both snubbing Lidar based syste

by Phlarp 7y ago

I agree in a big way here, I have no particular speciality with autonomous driving but I have seen enough media coverage of Musk both snubbing Lidar based systems and then Tesla auto-pilot hitting overhanging rear loads, which a lidar would seem to help with if not completely fix. His aversion to Lidar seems to have more to do with the cost implications on a vehicle that doesn't have a lot of room to raise the price or cut into the margins.

Someone1234 7y ago
Even parallax based systems are good at avoiding stationary objects, and are inexpensive since they use two overlapping visible light cameras (see EyeSight for one example). Essentially they work like the human eye to determine depth/distance. Tesla put two visible light cameras in their cars, but didn't configure them to parallax and instead rely on radar that has noise problems and ultrasound which only works at close range. It was a mistake that has had consequences.
- cr0sh 7y ago
  > Even parallax based systems are good at avoiding stationary objects, and are inexpensive since they use two overlapping visible light cameras (see EyeSight for one example). Essentially they work like the human eye to determine depth/distance. While I agree that it would be better to be using a multi-camera system for depth estimation, and there's no reason not to (because cameras are cheap and easy to build into bodywork of a vehicle), it shouldn't necessarily be needed. Otherwise, you'd have one-eye sighted people, who drive, running into the same obstacles, right? But they don't - so how do they do it? Probably motion cues, depth-from-motion, etc - something the brain does. From a cost-savings perspective, eliminating that extra camera will likely be looked at being done at some point, because over millions of cars manufactured, not installing that extra camera and wiring will save a ton of money and increase profitability. But Tesla isn't quite there yet, and probably shouldn't eliminate it... Furthermore, I'm not sure even LIDAR should be eliminated; the more sensors and sensor types, the better, ultimately. The problem is finding a LIDAR unit that has enough resolution and can be fitted to a vehicle (and doesn't look like a giant spinning top). Flash LIDAR units will probably be the solution, but they are still fairly expensive, and I don't know if they have the resolution needed yet, either.
  - MaulingMonkey 7y ago
    Some people do end up having trouble judging distance. Also a head can move around side to side, giving parallax over time even with one input (something you especially see birds of prey taking advantage of, to compensate for their narrower heads), whereas most cameras are on fixed mounts. Also, millions of years of evolution plus typically decades of training (in recognizing objects and distances) for the typical driver has a bit of an advantage over current neural nets, currently, for image recognition. It wouldn't suprise me if we crack that problem within my lifetime, but it also wouldn't suprise me if it takes awhile.
  - cstejerean 7y ago
    People with one eye do struggle with depth perception relative to people with two good eyes. It’s not completely gone but for example fast moving objects coming at you can be tricky (like catching a ball) as are point light sources at night (like figuring out exactly how far stop lights are). Speaking from experience that is, not sure how this generalizes across the population.
- nick_kline 7y ago
  Yes, you'd think adding more sensors would be better. However, even systems with lidar have run into vehicles and people. Google's system in cali ran into the bus, and then the uber system in Phoenix hit a person. Tesla's system has massively more users than all the other systems. They should get some benefit out of that data. I do worry that telsa will fail to make the software work well enough and get sued by all their existing customers (I'm a customer, but I want it to work :-)).
option 7y ago
Two things: 1) humans do it by vision alone, hence, it is possible in principle, and 2) lidars are expensive and introduce one more failure point. Most of the self-driving cars you see in bay area with lidars on top actually use those lidars to produce labeled data for vision-based distance estimator networks

by Phlarp 7y ago

Someone1234 7y ago

Even parallax based systems are good at avoiding stationary objects, and are inexpensive since they use two overlapping visible light cameras (see EyeSight for one example). Essentially they work like the human eye to determine depth/distance. Tesla put two visible light cameras in their cars, but didn't configure them to parallax and instead rely on radar that has noise problems and ultrasound which only works at close range. It was a mistake that has had consequences.

cr0sh 7y ago
> Even parallax based systems are good at avoiding stationary objects, and are inexpensive since they use two overlapping visible light cameras (see EyeSight for one example). Essentially they work like the human eye to determine depth/distance. While I agree that it would be better to be using a multi-camera system for depth estimation, and there's no reason not to (because cameras are cheap and easy to build into bodywork of a vehicle), it shouldn't necessarily be needed. Otherwise, you'd have one-eye sighted people, who drive, running into the same obstacles, right? But they don't - so how do they do it? Probably motion cues, depth-from-motion, etc - something the brain does. From a cost-savings perspective, eliminating that extra camera will likely be looked at being done at some point, because over millions of cars manufactured, not installing that extra camera and wiring will save a ton of money and increase profitability. But Tesla isn't quite there yet, and probably shouldn't eliminate it... Furthermore, I'm not sure even LIDAR should be eliminated; the more sensors and sensor types, the better, ultimately. The problem is finding a LIDAR unit that has enough resolution and can be fitted to a vehicle (and doesn't look like a giant spinning top). Flash LIDAR units will probably be the solution, but they are still fairly expensive, and I don't know if they have the resolution needed yet, either.
- MaulingMonkey 7y ago
  Some people do end up having trouble judging distance. Also a head can move around side to side, giving parallax over time even with one input (something you especially see birds of prey taking advantage of, to compensate for their narrower heads), whereas most cameras are on fixed mounts. Also, millions of years of evolution plus typically decades of training (in recognizing objects and distances) for the typical driver has a bit of an advantage over current neural nets, currently, for image recognition. It wouldn't suprise me if we crack that problem within my lifetime, but it also wouldn't suprise me if it takes awhile.
- cstejerean 7y ago
  People with one eye do struggle with depth perception relative to people with two good eyes. It’s not completely gone but for example fast moving objects coming at you can be tricky (like catching a ball) as are point light sources at night (like figuring out exactly how far stop lights are). Speaking from experience that is, not sure how this generalizes across the population.
nick_kline 7y ago
Yes, you'd think adding more sensors would be better. However, even systems with lidar have run into vehicles and people. Google's system in cali ran into the bus, and then the uber system in Phoenix hit a person. Tesla's system has massively more users than all the other systems. They should get some benefit out of that data. I do worry that telsa will fail to make the software work well enough and get sued by all their existing customers (I'm a customer, but I want it to work :-)).

option 7y ago

Two things: 1) humans do it by vision alone, hence, it is possible in principle, and 2) lidars are expensive and introduce one more failure point. Most of the self-driving cars you see in bay area with lidars on top actually use those lidars to produce labeled data for vision-based distance estimator networks