Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Cornell University Study on Cameras vs LiDAR

This site may earn commission on affiliate links.

SMAlset

Well-Known Member
Mar 4, 2017
9,437
10,378
SF Bay Area
Saw an article yesterday on Gizmodo about whether Cornell University’s study on using two stereo cameras up front on the windshield was a better idea and less expensive one than using a LiDAR system and whether this showed that Tesla was correct in its approach (although not much on this directly in the article). https://gizmodo.com/elon-musk-was-right-cheap-cameras-could-replace-lidar-1834266742

As you know Tesla has 3 cameras in use in the windshield rear view camera mounting.

Here’s a link to the Cornell University paper to be presented in June at a conference:

https://arxiv.org/pdf/1812.07179.pdf

The paper or the press on it makes it sound like this approach is new but isn’t this basically what Tesla has been doing all along? Maybe this is “new” research to some out there? Glad to see the study but confused.
 
Last edited:
If other research is done validating this paper (since no one wants to seemingly believe Tesla’s approach), I would think it will affect companies making LiDAR sensors and companies using them in their FSD technology approach. Surely some big bucks invested in the sensors and lots of research time. Curious to see where this goes.
 
Here's the bottom line for the paper: "In this paper we provide an alternative explanation with significant performance implications. We posit that the major cause for the performance gap between stereo and LiDAR is not a discrepancy in depth accuracy, but a poor choice of representation of the 3D information for ConvNetbased 3D object detection systems operating on stereo".

In plain english, what they're saying is that the reason LIDAR is considered more accurate is not because it is actually more accurate, but rather because the LIDAR-based systems have better object detection algorithms than the ones that integrate with optical cameras.

What they basically did is they took data from some optical cameras, turned it into a depth map and then used this depth map to generate a signal that matches the format that LIDAR systems generate.

They could then plug this signal into pre-existing object detection systems designed to work with LIDAR and they got much better results than the systems that were built directly on the camera signal were getting.

This is why they call their solution "pseudo-LIDAR" - it's a system that uses regular cameras to generate a signal that's similar to what LIDAR generates.

What it basically says is that if you use the right algorithms, cameras should do just fine.
 
Perhaps LIDAR is ultimately proven to be better than camera data. In its present form factor however, as used on current and projected vehicles (Waymo, Cruise, etc), for "taxi" use, for a consumer car for purchase or lease LIDAR won't be viable due to the roof mounted hardware that destroys the esthetic of the vehicle. Who will want to buy a car with a twirling instruments on a rack on top no matter how well it may be visually integrated? I think that is the main reason why Tesla is shunning LIDAR aside from its cost.