4 ms·

There are surprisingly sophisticated (and precise) machine learning algorithms for mono-camera depth estimation. It's not a problem anymore. Furthermore, most c

There are surprisingly sophisticated (and precise) machine learning algorithms for mono-camera depth estimation. It's not a problem anymore. Furthermore, most companies use various focal length lenses. The real problem lies in sensor fusion. Which sensor's input do you believe most, especially if one or more is occluded/dirty.

amelius 6y ago
I have yet to see any ML algorithm that can reach 99.999% accuracy or better (we need better).
- evdoks 6y ago
  Why do we even need such accuracy? Humans are extremely bad at estimating distance and nevertheless are quite ok drivers. ML needs to make sure that the car is not bumping at things or people, which is a very different task.
  - amelius 6y ago
    Yes, so "can/cannot bump into" is a classification tasks which needs quite some accuracy. A human on the other hand can recognize another human with near perfect accuracy.
  - Spivak 6y ago
    Because the failure modes are different. Humans with basically 100% accuracy both correctly recognize depth and put it into the right bucket of near/far but are woefully bad at estimating the numeric distance. So you’re right that you don’t need to do better than extremely coarse estimates like close, a few car lengths, in the distance, etc.. If computer vision fails at all to recognize depth (semi truck with clouds painted on it isn’t recognized) or misclassifies whether something is near or far that’s much more dangerous.
KaiserPro 6y ago
> There are surprisingly sophisticated (and precise) machine learning algorithms for mono-camera depth estimation There are, and I wouldn't trust them with life critical stuff. The more accurate ones are not realtime. The hardpart is that they are very noisy. unknown objects wobble about in depth considerably. You need to do lots of filtering to get useful results, which eats into time. It is very much an unsolved problem. Its one that I'm partly working on now. However for the device that I"m partly working with, monocular estimation is far too power hungry, noisy and generally shite. > Furthermore, most companies use various focal length lenses. no, tesla _rely_ on having a wide, medium and zoom camera. However they blended into the same effective sensor to give a better chance at tracking objects. I bet you a $10 that if you block one, the whole system turns to shit. > The real problem lies in sensor fusion. Sensor fusion is trivial. Its understanding what they sensors are telling you, thats the hard part. for example, when trying to turn across oncoming traffic, all the sensors will tell you that stuff is coming towards you, what they won't tell you is if its safe to cross. Thats the hard part. Given that tesla can't accurately place a car on the road yet, they can't safely cross traffic. Sensor understanding is the problem that is really unsolved. Depth estimation with stereo cameras is 85% of the way there, monocular estimation is no where near that level, simply because robust object recognition isn't going to be a thing for at least another 6 years.

amelius 6y ago

I have yet to see any ML algorithm that can reach 99.999% accuracy or better (we need better).

evdoks 6y ago
Why do we even need such accuracy? Humans are extremely bad at estimating distance and nevertheless are quite ok drivers. ML needs to make sure that the car is not bumping at things or people, which is a very different task.
- amelius 6y ago
  Yes, so "can/cannot bump into" is a classification tasks which needs quite some accuracy. A human on the other hand can recognize another human with near perfect accuracy.
- Spivak 6y ago
  Because the failure modes are different. Humans with basically 100% accuracy both correctly recognize depth and put it into the right bucket of near/far but are woefully bad at estimating the numeric distance. So you’re right that you don’t need to do better than extremely coarse estimates like close, a few car lengths, in the distance, etc.. If computer vision fails at all to recognize depth (semi truck with clouds painted on it isn’t recognized) or misclassifies whether something is near or far that’s much more dangerous.

KaiserPro 6y ago

> There are surprisingly sophisticated (and precise) machine learning algorithms for mono-camera depth estimation There are, and I wouldn't trust them with life critical stuff. The more accurate ones are not realtime. The hardpart is that they are very noisy. unknown objects wobble about in depth considerably. You need to do lots of filtering to get useful results, which eats into time. It is very much an unsolved problem. Its one that I'm partly working on now. However for the device that I"m partly working with, monocular estimation is far too power hungry, noisy and generally shite. > Furthermore, most companies use various focal length lenses. no, tesla _rely_ on having a wide, medium and zoom camera. However they blended into the same effective sensor to give a better chance at tracking objects. I bet you a $10 that if you block one, the whole system turns to shit. > The real problem lies in sensor fusion. Sensor fusion is trivial. Its understanding what they sensors are telling you, thats the hard part. for example, when trying to turn across oncoming traffic, all the sensors will tell you that stuff is coming towards you, what they won't tell you is if its safe to cross. Thats the hard part. Given that tesla can't accurately place a car on the road yet, they can't safely cross traffic. Sensor understanding is the problem that is really unsolved. Depth estimation with stereo cameras is 85% of the way there, monocular estimation is no where near that level, simply because robust object recognition isn't going to be a thing for at least another 6 years.