Is SLAM using cameras and HD maps (but no lidar) a solved problem for self-driving cars?

[strangecosmos]

I’ve heard oddly conflicting things about this from different sources, so I posed the question on Quora:

Is SLAM using cameras and HD maps (but no LIDAR), a solved problem for self-driving cars?

SLAM stands for Simultaneous Localization And Mapping.

Some academic research (one example) claims that camera-based localization without HD maps is sufficiently accurate for autonomous driving. Mobileye claims it can do extremely accurate localization with cameras and HD maps (which are compiled by cameras).

Yet technology analyst Ben Evans writes that camera-based SLAM doesn’t work yet, and Brett Winton, another technology analyst, agrees.

I hope to hear from a computer vision and/or autonomous driving expert on Quora. Can anyone here cite expert opinion on the matter?

 

[strangecosmos]

Brad Templeton answered on Quora, and Zeeshan Zia commented on his answer.

 

[strangecosmos]

Zeeshan Zia says it’s not a solved problem.

 

[strangecosmos]

Trying to get clarity on whether SLAM/odometry is solved using HD maps and stereo vision with a binocular or trinocular camera set up. (Mobileye and Tesla are both using a similar trinocular set up.) A lot of the research papers are about monocular vision, which is kind of funny. Why so much effort on monocular cameras if adding a second or third camera makes the problem easier?

I also realized I actually don’t know if SLAM/odometry is even solved using HD maps and lidar plus cameras. Anyone know?

 

[Bladerskb]

Mobileye doesnt use SLAM. What it uses (object detection using eyeq4) to build its maps is as accurate as a $10,000 IMU, his words not mine. Plus REM MAPPING is already in production TODAY. So were not talking about research here. We are not guessing whether something is solved or not. It is. It's in production

Research people are clueless when it actually comes to bleeding edge technology. A Research paper aims to hit a halfcourt shot by shooting once. A production developed system. Aims to hit the halfcourt shot anyway it can. That includes shooting, missing, grabbing the rebound and shooting from where it got the ball until it ends up close enough for a layup.

Published research only concentrates on novelty and therefore is not a viable indication of the status of actual production R&D!

 

[strangecosmos]

Mobileye doesnt use SLAM.

Mobileye has to do localization, if not Simultaneous Localization And Mapping (SLAM). That is, the car has to know where it is, and where other objects are in relation to the car.

 

[Bladerskb]

Mobileye has to do localization, if not Simultaneous Localization And Mapping (SLAM). That is, the car has to know where it is, and where other objects are in relation to the car.

Yeah they are doing localization but they are not using SLAM though. Which is synonymous with 3d point cloud today.

While the name sounds generic, it really isn't. visual SLAM is very specific to tracking a set of points through successive camera frames, and using these tracks to triangulate their 3D position; while simultaneously using the estimated point locations to calculate the camera pose which could have observed them.

What Mobileye is doing is completely different and can't be called SLAM which is why they don't call it SLAM.

 

[strangecosmos]

What is Mobileye doing differently?

 

[Bladerskb]

What is Mobileye doing differently?

This is probably better, watch from 41:20, he mentions SLAM at 48 mins

"Its a variation on SLAM but its a bit more clever cause SLAM is heavy in terms of data"

 

[strangecosmos]

What he’s talking about at 48 minutes in that video when he mentions SLAM is compiling HD maps from camera images — mapping, not localization. A few minutes later he says that Mobileye vehicles do localization with the help of HD maps.