Will Tesla ever do LIDAR?

[StealthP3D]

As long as what you care about is in at least two sufficiently distinct frames —either from different angles or different zoom settings — you can compute a depth map using only procedural code, without any neural network involvement.

Now that's what I'm talking about!

Except for one small correction. A depth map can only be created from two frames if the lenses were in a different position from each other. A different zoom setting on the same lens, in the same position, is not sufficient. Two consecutive frames might be sufficient if the lenses were in a moving car (because then the lens would be in a different position for each frame). However, the depth map will be better for most purposes if the two images are taken simultaneously from slightly different locations because then you are not complicating things by introducing the problem of other objects moving in the scene (between the two consecutive images).

All of this is made possible by the explosion of cheap computing power over the last several decades.

 

[heltok]

An interesting case where both Camera and Lidar will struggle =)

 

[StealthP3D]

An interesting case where both Camera and Lidar will struggle =)

I doubt we will be seeing any roadside objects coated in Vantablack! However, a person standing in the middle of the road at the top of a rise against a sky background, would be perfectly visible to a camera (as a silhouette of a person) and completely invisible to LIDAR because LIDAR is handicapped by needing to get a return reflection of its laser beam while a camera can detect the brightness difference between the object and the sky background. Even black objects give LIDAR a real challenge at a distance.

 

[CarlK]

Now that's what I'm talking about!

Except for one small correction. A depth map can only be created from two frames if the lenses were in a different position from each other. A different zoom setting on the same lens, in the same position, is not sufficient. Two consecutive frames might be sufficient if the lenses were in a moving car (because then the lens would be in a different position for each frame). However, the depth map will be better for most purposes if the two images are taken simultaneously from slightly different locations because then you are not complicating things by introducing the problem of other objects moving in the scene (between the two consecutive images).

All of this is made possible by the explosion of cheap computing power over the last several decades.

You can actually still get depth info even with a single frame from one lens by using two sensors to measure difference between lights coming from two extreme edges of the lens. That's how phase detection auto focus in camera works. Two lenses separated by a distance is much better of course.

 

[Shygar]

How on earth do humans drive without built in lidar and a 20 watt processor

 

[JHWJR]

How many more fire trucks must be injured until people will realize that maybe this whole radar+camera isn't working?

You think the issue with "firetrucks" is because it can't see the firetruck and LIDAR will help? If so, you are lost.

 

[dgatwood]

Now that's what I'm talking about!

Except for one small correction. A depth map can only be created from two frames if the lenses were in a different position from each other. A different zoom setting on the same lens, in the same position, is not sufficient.

Are you sure? I could have sworn I remember reading something about taking advantage of differences in the way lenses focus light at different zoom settings to determine distance. *shrugs*

You can actually still get depth info even with a single frame from one lens by using two sensors to measure difference between lights coming from two extreme edges of the lens. That's how phase detection auto focus in camera works. Two lenses separated by a distance is much better of course.

You can also can generate a depth map with two images from the same camera using different aperture settings, even with everything else equal, though this probably isn't practical to do with moving video.

Either way, parallax is by far the simplest approach.

 

[S4WRXTTCS]

You think the issue with "firetrucks" is because it can't see the firetruck and LIDAR will help? If so, you are lost.

It was a clearly exaggerated example of the struggle radar has with stopped objects, and the struggles Tesla has had in incorporating vision.

I do expect to see a lot of improvements with Tesla vision especially with HW3.

But, it's impossible to say if those improvements will be enough to achieve the level of safety and redundancy really required for autonomous driving at L3 or above.

 

[S4WRXTTCS]

How on earth do humans drive without built in lidar and a 20 watt processor

I can't speak to anyone else, but my sixth sense has saved me a few times.

I'm just trying to be a robot advocate.

Robots want lidar.

 

[StealthP3D]

I can't speak to anyone else, but my sixth sense has saved me a few times.

What sense is "the sixth sense"? Oh, that's right, it's magic.

The real problem with human drivers is they are inconsistent, distracted and make unwarranted assumptions.

 

[S4WRXTTCS]

What sense is "the sixth sense"? Oh, that's right, it's magic.

The real problem with human drivers is they are inconsistent, distracted and make unwarranted assumptions.

I don't think there is anything magical about it. For me it's probably the subconscious matching a pattern to something it's seen before. Just because I don't understand exactly how it works doesn't make it magical. Human drivers are indeed all of that, and a lot of them are on drugs. There is noting 6th sense about being extra vigilant while driving near a pot dispensary.

What humans have though is flexibility.

We're allowed to fail which autonomous cars largely aren't. This means our sensors can be imperfect (up to an amount), and we don't require redundancy.

We're allowed to be unpredictable, and in fact we use this unpredictability to keep us safe from being bullied on the road.

We're not just allowed to bend the written rules of the road, but it's expected.

It's much easier to talk about the pro's, and con's of various sensor technologies (Radar, Lidar, TOF Sensors (this includes stereo cameras). etc) than it is about how to exactly get autonomous cars and human driven cars to coexist on our roads.

In any case the actual answer to the persons question is that autonomous cars have a requirement for redundancy, and that's not required from human drivers.

This of course leads to the question of whether redundancy really should be a hard requirement.

 

[StealthP3D]

I don't think there is anything magical about it. For me it's probably the subconscious matching a pattern to something it's seen before.

If it's just pattern matching, then it's not a sixth sense and computers are actually more suited to make good use of that than humans.

By far.

There is noting 6th sense about being extra vigilant while driving near a pot dispensary.

You bring up another advantage of computers - they are never NOT "extra vigilant". They don't let their guard down or get lazy like humans.

What humans have though is flexibility.

I hope you're not presenting that as an advantage!

We're allowed to fail which autonomous cars largely aren't. This means our sensors can be imperfect (up to an amount), and we don't require redundancy.

If redundancy is built into the FSD system, that's a plus. I'm not sure how the fact that humans don't have it can be considered an advantage. Redundancy is a GOOD thing! And I disagree that autonomous cars are not allowed to fail. They are expected to fail occasionally. But as long as it's at a lower rate than humans fail, it's a net benefit.

That reminds me of an old quote:

"I would rather die peacefully in my sleep like my grandfather rather than screaming for their lives like his passengers."

 

[S4WRXTTCS]

And I disagree that autonomous cars are not allowed to fail. They are expected to fail occasionally. But as long as it's at a lower rate than humans fail, it's a net benefit.

It's not that they won't be allowed to fail, but the question of what sort of failure rate would be allowed.

You say a lower rate, but what rate?

I would say that most human drivers are better than "average" human drivers because a few bad apples cause a lot of the accidents, and issues on our roads.

Repeated DUI offenders
Habitual speeders (well beyond the speed limit)
Inexperienced drivers

Then there are behavior things that change. It appeared like fatalities related to distracted driving would continue to go up, and in fact I argued that autonomous cars were inevitable because of how crappy human drivers were getting. But, it looks like we already reached peak distraction. But, that might just be an anomaly.

Autonomous cars are the ultimate trolley car problem.

To save people you have to kill some.

 

[StealthP3D]

It's not that they won't be allowed to fail, but the question of what sort of failure rate would be allowed.

You say a lower rate, but what rate?

I would say that most human drivers are better than "average" human drivers because a few bad apples cause a lot of the accidents, and issues on our roads.

Repeated DUI offenders
Habitual speeders (well beyond the speed limit)
Inexperienced drivers

Then there are behavior things that change. It appeared like fatalities related to distracted driving would continue to go up, and in fact I argued that autonomous cars were inevitable because of how crappy human drivers were getting. But, it looks like we already reached peak distraction. But, that might just be an anomaly.

I've got news for you. Humans have always been terrible drivers.

To save people you have to kill some.

That's a very pessimistic way to look at it. I would say "To save lives you need to reduce the slaughter rate on our highways".

While it's always good to make our roads safer just for the death and misery it will avoid, there is also the economic benefit which will be huge.

 

[StealthP3D]

A British startup has a strong idea that by using AI, LIDAR is not necessary and all you need is a monocular forward facing camera:

British startup says its self-driving car doesn't need fancy tech

Watch the video too.

 

[electronblue]

A British startup has a strong idea that by using AI, LIDAR is not necessary and all you need is a monocular forward facing camera:

British startup says its self-driving car doesn't need fancy tech

Watch the video too.

The problem with startups is you never know when they are right or when they just have a bold idea that will eventually fail. That is the whole premise of investing in startup stories of course. The whole idea behind venture capitalism.

Incidentally we were discussing this fake it till you make it culture in another thread. The problem is you never know who is actually going to make it and who is Theranos.

 

[StealthP3D]

The problem is you never know who is actually going to make it and who is Theranos.

I've been using AP in our LR Model 3 for one year next month and the improvements I've seen in less than one year are nothing short of stunning. If this pace keeps up, I have no doubt the system will be very capable within 2 years.

But I like hearing the position of the naysayers - it reminds me of those who thought the notion of a man visiting the moon to be utterly ridiculous, preposterous and impossible. Some of them believed this so stubbornly that once it was accomplished, less than 10 years later, they refused to believe it was real. They went to their graves believing the whole thing was a hoax!

Because no human could achieve preposterous and impossible things.

 

[electronblue]

I've been using AP in our LR Model 3 for one year next month and the improvements I've seen in less than one year are nothing short of stunning. If this pace keeps up, I have no doubt the system will be very capable within 2 years.

But I like hearing the position of the naysayers - it reminds me of those who thought the notion of a man visiting the moon to be utterly ridiculous, preposterous and impossible. Some of them believed this so stubbornly that once it was accomplished, less than 10 years later, they refused to believe it was real. They went to their graves believing the whole thing was a hoax!

Because no human could achieve preposterous and impossible things.

That’s the thing isn’t it.

You see me as a naysayer where as I see you as a Lidar naysayer.

I don’t think either one of us intends to be any such thing though. I believe I am keeping an open mind, both critically and optimistically too.

Still, some companies indeed are Theranos — and even they may have meant well.

 

[CarlK]

I've got news for you. Humans have always been terrible drivers.

That's very true. Many people are so self-absorded they have lost objectivity. The same as many people's opinion that AI will not exceed human intelligence. The fact is human intelligence is just an arbitrary point on the evolution scale. There is nothing special of that.

That's a very pessimistic way to look at it. I would say "To save lives you need to reduce the slaughter rate on our highways".

While it's always good to make our roads safer just for the death and misery it will avoid, there is also the economic benefit which will be huge.

The good news is there are still rational people out there. When autonomous cars can be proven to be several times safer than human drivers driving your own car will not be affordable except for a very few, if it's not outright banned.

 

[StealthP3D]

That’s the thing isn’t it.

You see me as a naysayer where as I see you as a Lidar naysayer.

I believe LIDAR is not a necessary component of self-driving cars. If that makes me a "LIDAR naysayer", fair enough. But I'll point out that I'm not saying LIDAR doesn't work, I think you could make self-driving work with ONLY LIDAR (just don't drive in the rain, snow or fog). And I don't think it would be as safe as a system based on cameras which are capable of providing more useful information.

I don’t think either one of us intends to be any such thing though. I believe I am keeping an open mind, both critically and optimistically too.

I guess if you can't tell whether Elon Musk is another Theranos or not, then your mind is more "open" than mine (but not in an "optimistic way). Because I've already concluded that he is not based on the fact that I have two of his products and they both work as advertised. Theranos was a massive fraud and Elizabeth Holmes a massive fraudster with products that worked so poorly they were secretly using standard blood testing machines to process their blood tests. But Tesla's product is right in front of me to use and enjoy. And it keeps getting better at a blindingly fast rate (I've only had a Model 3 for less than a year and the AP is already leaps and bounds better than when I took delivery).