Will Tesla ever do LIDAR?

[diplomat33]

Here is article in Cornell Chronicle also from last Monday where they are finding the benefits in camera vision vs LIDAR, too, and confirm the massive cost difference.
New way to ‘see’ objects accelerates future of self-driving cars | Cornell Chronicle
They will present their paper in Long Beach in June

Thanks. Looks like Musk was right.

 

[Daniel in SD]

Thanks. Looks like Musk was right.

This is the same paper that was posted before. They said it was a solid starting point. I don’t think anyone is denying that it’s possible to get depth data from cameras.

Table 4 shows our results with F-POINTNET and com-
pares to those with LiDAR, on the validation set. Compared to the car category (cf. Table 1), the performance gap is significant. We also observe a similar trend that the gap be- comes larger when moving to the hard cases. Nevertheless, our approach has set a solid starting point for image-based pedestrian and cyclist detection for future work.

 

[StealthP3D]

This is the same paper that was posted before. They said it was a solid starting point. I don’t think anyone is denying that it’s possible to get depth data from cameras.

Actually, that's a common claim I've seen from neophytes who argue LIDAR is a necessary component of autonomy. That Tesla's efforts are misguided because they lack the accurate depth information that LIDAR provides. Mostly they are just arguing in an attempt to sow FUD (without really knowing what they are talking about).

 

[heltok]

It should be noted that the paper used low res images (0.4Mp) while it seems that Tesla’s AKNET_V9 uses the full res of the camera(not sure but think it is 1Mp?!). The higher resolution in combination with using more cameras(of which one is narrow) and Radar should improved hard cases(long distance) of object detection greatly.

 

[CarlK]

Have we already settled this? Lidar is useless. It can't function as a primary sensor and it's unnecessary and not the best choice as a secondary sensor. It makes neither technological nor economical sense. Let's hear it if anyone can think of a (fact based) justification for it.

 

[StealthP3D]

Let's hear it if anyone can think of a (fact based) justification for it.

Fact #1: Elon Musk is a liar and a fraudster.
Fact #2: Elon Musk is not using LIDAR.
Conclusion: LIDAR is necessary for autonomy to work.

 

[electronblue]

Fact #1: Elon Musk is a liar and a fraudster.
Fact #2: Elon Musk is not using LIDAR.
Conclusion: LIDAR is necessary for autonomy to work.

If only logic was that simple...

I don’t think there is any reason to suggest any particular technology is absolutely necessary for autonomy to work. For all we know some new kind of sensor appears tomorrow that puts everything else into shame.

What the constructive discussion in my view mostly revolves around are two things:

1) What kind of minimum sensor suite is necessary for autonomy to work at a sufficient safety level for true car responsible driving?

2) What is the sensor suite in the car that first ”makes it big” with true autonomous (car responsible) driving?

The answer to these questions may or may not be the same. It may also be that we haven’t seen the answer to either of these questions yet in any car. The question, for me at least, then isn’t about ”is Lidar absolutely necessary”, but about ”do I believe someone using Lidar might have better answers to 1) and 2) than someone not using Lidar”? And to that I can’t conclusively answer no so the questions remain open and Tesla’s claim remains unproven.

 

[electronblue]

Tesla’s claims of course go even further. On the 22nd they claimed there was no reason for anyone to want Lidar in their car.

That is an even bolder claim than the claim that Lidar is not needed for autonomous driving.

It suggests a car with Lidar could not for example be any safer than a car without. I think that claim is even harder to digest at this stage.

Because if it comes to your children on the autonomous school run being more safe by addition of Lidar, there will be a reason to have Lidar and people will want Lidar. Heck they will want the biggest dome on the roof if it helps.

 

[iTech]

Have we already settled this? Lidar is useless. It can't function as a primary sensor and it's unnecessary and not the best choice as a secondary sensor. It makes neither technological nor economical sense. Let's hear it if anyone can think of a (fact based) justification for it.

Have you considered competing in the 2020 Summer Olympics? You seem to be very good at gymnastics.

 

[CarlK]

Have you considered competing in the 2020 Summer Olympics? You seem to be very good at gymnastics.

ITech got you beat.

 

[StealthP3D]

Let's hear it if anyone can think of a (fact based) justification for it.

So far, I'm hearing crickets. Not surprising.

 

[electronblue]

Have we already settled this? Lidar is useless. It can't function as a primary sensor and it's unnecessary and not the best choice as a secondary sensor. It makes neither technological nor economical sense. Let's hear it if anyone can think of a (fact based) justification for it.

I already gave my more complex answer above in #147 but to answer your question on justification for Lidar:

It has practically zero false negatives — it is inherently better than radar, ultrasonics or vision at seeing ”something is there and it is this and his far away”. That alone makes it a great crash-detection sensor — and indeed a great redundancy sensor.

Also Lidar does this in complete darkness.

I think people are rushing to judgement on the optimal sensor suite based on what or who they are fans of. Nobody knows the optimal suite yet.

 

[StealthP3D]

I already gave my more complex answer above in #147 but to answer your question on justification for Lidar:

It has practically zero false negatives — it is inherently better than radar, ultrasonics or vision at seeing ”something is there and it is this and his far away”.

Actually it doesn't "see something is there" better because it's notoriously poor at being able to tell a puff of smoke/ steam or a plastic grocery bag from a solid object - that makes it unreliable and dangerous to rely on. It also cannot see very far down the road, particularly black objects at a distance. Camera vision is much better at these things.

Also Lidar does this in complete darkness.

How much driving will be occurring in complete darkness when you agree that LIDAR is only supplemental to primary camera vision? Human's don't drive in complete darkness either - they use something called headlights. Driving in complete darkness simply doesn't happen because the road is lit (either by streetlights and/or headlights). How is working in complete darkness an advantage when you acknowledge it can only be used in conjunction with camera vision?

CarlK asked for FACT based justification. Simply saying it works better is not fact-based.

 

[Daniel in SD]

Have we already settled this? Lidar is useless. It can't function as a primary sensor and it's unnecessary and not the best choice as a secondary sensor. It makes neither technological nor economical sense. Let's hear it if anyone can think of a (fact based) justification for it.

Every single autopilot accident I’ve seen a video of could have been avoided by a system using LIDAR.
The paper you all keep citing says that LIDAR works better than vision right now. Nobody know what is necessary for FSD. Tesla mentioned during their presentation that they’re working on an even faster NN chip. Maybe HW3 will be good enough (I doubt it) or maybe HW10 will be good enough or maybe there will never be FSD.

 

[electronblue]

Actually it doesn't "see something is there" better because it's notoriously poor at being able to tell a puff of smoke/ steam or a plastic grocery bag from a solid object - that makes it unreliable and dangerous to rely on. It also cannot see very far down the road, particularly black objects at a distance. Camera vision is much better at these things.

I never claimed Lidar was necessarily great at avoiding false positives. I said it is great at avoiding false negatives and ranging those hits. It is a unique feature where it excels over the other currently used sensor types. That can certainly make it useful and many seem to think it does.

As for camera being better that comes back to training the vision — and that can happen to both cameras and Lidar inputs.

How much driving will be occurring in complete darkness when you agree that LIDAR is only supplemental to primary camera vision? Human's don't drive in complete darkness either - they use something called headlights. Driving in complete darkness simply doesn't happen because the road is lit (either by streetlights and/or headlights). How is working in complete darkness an advantage when you acknowledge it can only be used in conjunction with camera vision?

Easy. Seeing that headlight-less car or animal approaching from the side while your headlights are pointed in the other direction — again it comes back to vision on those 90 degree turns. With Tesla it is all down to the B pillar camera...

CarlK asked for FACT based justification. Simply saying it works better is not fact-based.

Lidar’s features are fact just as much as anything else in this conversation.

 

[StealthP3D]

Lidar’s features are fact just as much as anything else in this conversation.

Vision features are fact too. You have still failed to use those facts to make a compelling reason why LIDAR is needed, how it should be incorporated and why humans are not required to use LIDAR before getting behind the wheel of a deadly machine on public roads.

Also, since you have such strong opinions on this subject, how many LIDAR do you recommend and where would they be mounted to be most effective?

 

[CarlK]

Every single autopilot accident I’ve seen a video of could have been avoided by a system using LIDAR.

Uber did not use Lidar?

The paper you all keep citing says that LIDAR works better than vision right now. Nobody know what is necessary for FSD. Tesla mentioned during their presentation that they’re working on an even faster NN chip. Maybe HW3 will be good enough (I doubt it) or maybe HW10 will be good enough or maybe there will never be FSD.

But what do you mean by "now"? If now is Tesla's HW3 and the pilot NN then nobody is even close. From what we understand now (BTW some experts probably already knew that) Lidar + HD mapping is a stop gap and quick fix that let you to put something out to impress investors, those who are putting money in your company or buying you out. It can get you to 98% or so relatively quickly but will never let you get the final one or two percent. You only realize that when you have reached 99%, like Lavendowski did, other than Elon who saw it from the beginning. You just have to give it to that guy.

Good enough may mean 2~3x better than human driver or even 10x better but it will never be really good enough. Things will not just stand still. That's the good thing but also the scary part.

A very strong tell tale sign is all those financial pundits are laughing at Elon for his comment but did you see even one Lidar guy, there are probably couple dozen companies that are involved in that, comes out to dispute that? Only Nvidia gave a pretty weak protest of the AI chip claim, it had to, but Lidar guys are unusually quite, even though what Elon said pretty much made their business worthless.

 

[Daniel in SD]

Vision features are fact too. You have still failed to use those facts to make a compelling reason why LIDAR is needed, how it should be incorporated and why humans are not required to use LIDAR before getting behind the wheel of a deadly machine on public roads.

Also, since you have such strong opinions on this subject, how many LIDAR do you recommend and where would they be mounted to be most effective?

No one is saying lidar will always be needed! All we’re saying is that it currently does certain things better than cameras right now. You also seem to be very confused about how lidar is used. It’s not a replacement for cameras, driving without cameras would be impossible.
Maybe in the future cars will be required to have lidar. Safety standards are always improving so I could see it happening. My guess is that cameras will be good enough for collision avoidance systems.

 

[Daniel in SD]

Uber did not use Lidar?

They didn’t use it. Read the report. Their camera/lidar system saw the pedestrian but they were trying to get their disengagement rate down so they ignored it.

But what do you mean by "now"? If now is Tesla's HW3 and the pilot NN then nobody is even close. From what we understand now (BTW some experts probably already knew that) Lidar + HD mapping is a stop gap and quick fix that let you to put something out to impress investors, those who are putting money in your company or buying you out. It can get you to 98% or so relatively quickly but will never let you get the final one or two percent. You only realize that when you have reached 99%, like Lavendowski did, other than Elon who saw it from the beginning. You just have to give it to that guy.

Everyone is using cameras! Lidar just provides more data that can not currently be obtained by cameras alone. There is zero evidence that Tesla has made their camera/radar system as good as others camera/radar/lidar systems.

 

[CarlK]

No one is saying lidar will always be needed! All we’re saying is that it currently does certain things better than cameras right now. You also seem to be very confused about how lidar is used. It’s not a replacement for cameras, driving without cameras would be impossible.
Maybe in the future cars will be required to have lidar. Safety standards are always improving so I could see it happening. My guess is that cameras will be good enough for collision avoidance systems.

You forget radar which is a much better complement to vision. Either way yours is a straw man argument. Nobody is using Lidar to complement vision. All those cars with big Lidar on top are using Lidar as a primary system. These two, Lidar + HD mapping and Vision + machine learning are two totally different technologies.