Tesla

[Morris Minor]

HAHA!!!

MOTHERFUCK YES

Musk said you can overpower & easily outrun them. But you know they're going to have to build an evil version.

[MrH]

https://venturebeat.com/2021/07/03/tesla-ai-chief-explains-why-self-driving-cars-dont-need-lidar/

Fact check: what they presented at AI day was largely done with LIDAR

[r0tor]

Fact check: what they presented at AI day was largely done with LIDAR

Now there is an amazing fact since the cars have no lidar!

[MrH]

Now there is an amazing fact since the cars have no lidar!

They have test vehicles with LIDAR.  That is what was used to develop a lot of the material for the presentation.

There are countless times in the presentation, where things that are clearly visibly blocked, that show up on their images.  Side streets that are blocked entirely by other builds, internal walls of buildings, etc.  Vision only systems can not generate that type of information.  There is also a clear canopy cut off, which is what happens with LIDAR.  Tree tops, etc are all clearly visible to front facing cams.

[Morris Minor]

Fact check: what they presented at AI day was largely done with LIDAR

That's right. You seem to think it could have taken them further?

[GoCougs]

Musk said you can overpower & easily outrun them. But you know they're going to have to build an evil version.

Sure, but that's not where my mind went  .

[r0tor]

They have test vehicles with LIDAR.  That is what was used to develop a lot of the material for the presentation.

There are countless times in the presentation, where things that are clearly visibly blocked, that show up on their images.  Side streets that are blocked entirely by other builds, internal walls of buildings, etc.  Vision only systems can not generate that type of information.  There is also a clear canopy cut off, which is what happens with LIDAR.  Tree tops, etc are all clearly visible to front facing cams.

You can find similar videos on YouTube from customer cars

I of course know you have some actual proof of your accusations - corrects?

[MrH]

You can find similar videos on YouTube from customer cars

I of course know you have some actual proof of your accusations - corrects?

There are literally interior walls, from a building around a corner.  You don't that with vision systems.

[r0tor]

 

That was a post processed video showing the accuracy of their AI vision system

[MrH]

???

The screen cap on the left isn't possible with vision in real time.  Did they drive the Tesla through walls to see into the buildings?  How did the vision system see the staircase going up?

[r0tor]

That video clip was a reconstructed model for validation purposes- not what the onboard system is processing.  Nothing about that video is "real time". 

A vision system is perfectly capable of looking through glass.  That 1 frame from that 1 camera may see a reflection, but with a slightly different angle from another camera or at a slightly later time perhaps would not see a reflection - or if lighting were to change.

[Morris Minor]

That video clip was a reconstructed model for validation purposes- not what the onboard system is processing.  Nothing about that video is "real time". 

A vision system is perfectly capable of looking through glass.  That 1 frame from that 1 camera may see a reflection, but with a slightly different angle from another camera or at a slightly later time perhaps would not see a reflection - or if lighting were to change.

I know AI day was mostly about recruiting AI engineering talent - & the robot is a moonshot sweetener to try & lure talent in. But I think it also served as a sandbag move: 100% of the MSM (& 95% of analysts) did not understand a word of what was going on. But they all picked up on the wackiness of  the humanoid robot and are writing about it. That's a useful smokescreen.

[r0tor]

Very true .. that event really speaks to the 1 in 100,000 people that can slightly understand what's going on

[GoCougs]

I'm gonna take a good guess at what's going on. FYI I use this tech but not at Waymo. In short, it is possible to get a 3D model from 2D imaging.

The first is stereoscopic vision - different cameras looking at the same feature/FOV - the slight difference in perspective is how a pseudo 3D model is created:



The second is photometic stereo - using shadowing from different perspectives of the same feature/FOV - the slight difference in shadowing is how a pseudo 3D model is created:



Both work well enough for industrial applications (i.e., robotics and such, which is my field) but it's not a 100% real representation. That's okay when the environment is highly consistent/controlled (i.e., predictable) as it is with a machine in a factory. A car driving down a road, in untold weather and light conditions, with untold road features and conditions, with untold variation with what is ON the road? Color my unsurprised that after all this time Tesla Autopilot still doesn't work even remotely close as promised.

[r0tor]

Tesla doesn't need to do a 1:1 vector graphic representation of the entire surrounding. They only need to detect things like cars, traffic signals, lane markers, curbs, pedestrians, ect... which is a lot easier if a task.

This is similar to photogrammetry (which is what your example is) but also very different.  The tech is more focused on object detection than physical representation.

[Morris Minor]

I'm gonna take a good guess at what's going on. FYI I use this tech but not at Waymo. In short, it is possible to get a 3D model from 2D imaging.

The first is stereoscopic vision - different cameras looking at the same feature/FOV - the slight difference in perspective is how a pseudo 3D model is created:



The second is photometic stereo - using shadowing from different perspectives of the same feature/FOV - the slight difference in shadowing is how a pseudo 3D model is created:



Both work well enough for industrial applications (i.e., robotics and such, which is my field) but it's not a 100% real representation. That's okay when the environment is highly consistent/controlled (i.e., predictable) as it is with a machine in a factory. A car driving down a road, in untold weather and light conditions, with untold road features and conditions, with untold variation with what is ON the road? Color my unsurprised that after all this time Tesla Autopilot still doesn't work even remotely close as promised.

12-minute summary of what they're doing & aiming for. The guy is a fan.
https://youtu.be/ABbDB6xri8o

[MrH]

Lex Fridman is a mess.

https://www.thedrive.com/tech/27359/mit-study-showing-high-functional-vigilance-among-autopilot-users-comes-with-massive-caveats

[MrH]

That video clip was a reconstructed model for validation purposes- not what the onboard system is processing.  Nothing about that video is "real time". 

A vision system is perfectly capable of looking through glass.  That 1 frame from that 1 camera may see a reflection, but with a slightly different angle from another camera or at a slightly later time perhaps would not see a reflection - or if lighting were to change.

So they're using the technology that Elon said is a "crutch" and Waymo is "friggin stupid" for using it, but yet Tesla is using it to try and validate their vision system?

Most of us insist on watching Netflix in 4k, but you think 1280x960 cameras are going to drive themselves.

[MrH]

Tesla doesn't need to do a 1:1 vector graphic representation of the entire surrounding. They only need to detect things like cars, traffic signals, lane markers, curbs, pedestrians, ect... which is a lot easier if a task.

This is similar to photogrammetry (which is what your example is) but also very different.  The tech is more focused on object detection than physical representation.

Very easy.  If only they could figure out the damn moon.

https://twitter.com/JordanTeslaTech/status/1418413307862585344?s=20

[Morris Minor]

Lex Fridman is a mess.

https://www.thedrive.com/tech/27359/mit-study-showing-high-functional-vigilance-among-autopilot-users-comes-with-massive-caveats

Why?

[Morris Minor]

Very easy.  If only they could figure out the damn moon.

https://twitter.com/JordanTeslaTech/status/1418413307862585344?s=20

Why wouldn't they add this to their training heap?

[MrH]

Why wouldn't they add this to their training heap?

I think this is where you and I aren't on the same page.  First: how do things get fed to the training heap?  When, how, and why? Second: how many things need to get fed to this training heap for this to be level 5 autonomy?  I'm curious what you will say for both.

[MrH]

Why?

Did you read the link?  That's exactly why.

[r0tor]

I think this is where you and I aren't on the same page.  First: how do things get fed to the training heap?  When, how, and why? Second: how many things need to get fed to this training heap for this to be level 5 autonomy?  I'm curious what you will say for both.

Your rant is why Tesla is by far and away in the lead on autonomous driving - the size and diversity of their training data lake.

[r0tor]

So they're using the technology that Elon said is a "crutch" and Waymo is "friggin stupid" for using it, but yet Tesla is using it to try and validate their vision system?

Most of us insist on watching Netflix in 4k, but you think 1280x960 cameras are going to drive themselves.

First, your assumption that the display was LiDAR is still just an assumption.  Second, using LiDAR as a validation parameter through post processing is a completely different ball game then trying to use it real time.

[MrH]

So you're just writing off what I post as an assumption, but can't explain to me how a vision system would map interior walls, staircases, buildings around the corner, etc.  It most definitely is lidar in those images.

[MrH]

Shot:




Chaser:

[Laconian]

He's pulling out all the stops! Time to order pizzas for the nerds and force them to work nights!

[Morris Minor]

Haha - someone on Reddit's all in:
(but no S3XY shorts or flamethrower)

https://www.reddit.com/r/Starlink/comments/pge1wa/ive_never_been_a_completionist_at_least_not_on/

[Morris Minor]

(would like to see more of the VW Kombi in the back)