Everything i talk about is real world deployment. All you are doing here is back peddling and trying to find excuses. What's it gonna be?
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Drago Anguelov – Machine Learning for Autonomous Driving at Scale
- Thread starter shrineofchance
- Start date
Everything i talk about is real world deployment. All you are doing here is back peddling and trying to find excuses. What's it gonna be?
powertoold
Active Member
No back peddling here, just making sure since you've historically had a loose definition of "do", "has", and "deployed".
To my knowledge, no one is using LIDAR to recognize road markings or signs as their goal solution for FSD. That isn't to say they haven't experimented with it. But there are thousands of failed / inadequate experiments that lead to an eventual solution.
DanCar
Active Member
List everything below that vision from camera does and then list everything that vision from lidar does and lets check.
For example, what below can lidar not do?
I'll get the ball rolling, feel free to correct me.
..................Lidar.........................Vision
Traffic Lights....Detect, not color.............Detect plus color
Signs
Overhead signs....Reflective + Detect...........Detect
Road Signs........Reflective + Detect...........Detect
Read Signs........Yes in theory. In Practice?...Yes
Markings
Crosswalks........Reflective Detect.............Detect
Road Markings ....Reflective Detect.............Detect
Lane Lines........Reflective Detect.............Detect
Objects
Curbs.............Detect........................Detect
Crosswalks........Detect........................Detect
Moving Objects....Detect plus velocity..........Detect
Static Objects....Detect........................Detect
powertoold
Active Member
This has got to be the biggest joke conversation I've had so far. To think lidar will be used to read signs or road markings for FSD.
Do you guys actually think LIDAR has enough resolution to recognize text on a road sign, let alone a road marking? Good God, bladerskb and his hundreds of academic papers.
Do you guys actually think LIDAR has enough resolution to recognize text on a road sign, let alone a road marking? Good God, bladerskb and his hundreds of academic papers.
Precisely
To the uninformed and misinformed, facts are regarded as a joke. You said "but lidar can't do the majority in that picture."
Why can lidar see lane lines, cross walk and road markings? Because of the luminosity and reflectance of the paint, secondly the raised bump of the lines. Infact Lidars can see lane markings better than they can see vehicles and pedestrains.
Here is Lidar from LIVOX detecting quite clearly:
Lane Lines, Cross Walk, Road Markings, Curbs, Overhead Signs, etc.
Here is LIDAR from Veloydyne clearly detecting & seeing:
Lane Markings, Lane Lines, Cross Walk Road Markings, Curbs, Construction, Overhead Signs, etc
Lidar from OUSTER clearly reading traffic/road signs (25 speed limit):
I could keep posting more from Luminar, Innuvision, Innoviz, but its pointless.
Most Tesla fan still think that Lidar Tech is stuck at where it was in 2010. Not knowing there has been over 10 years of Lidar development. Questions being asked in this thread is will companies drop lidar? What should be asked is will companies drop camera and only use it in front detect status of traffic lights?
Going even further than the above.
Modern Lidar can take the depth and intensity from each Lidar point returns to display fixed resolution depth images, signal images, and ambient images in real time, all without a camera. This isn't hooking up a camera next to a lidar. No, this is the actual output from the Lidar.
This is the output of Waymo's 5th gen Lidar.
This is the output of Outer's lidar after a firmware update in 2018 to use the Lidar point returns to display an image and as you can see the 50 speed limit sign:
Innoviz
Road Markings, Lane Markings, Lane Lines, Barrier, etc
Curb
Road Markings, Lane Markings, Lane Lines, Barrier, etc
Curb
powertoold
Active Member
Lol what a joke. You do know that detection is different than recognition. And then there's also % confidence?
Almost all fsd related discussions have a major logical flaw in that some people here think that being able to "do" or "detect" something means it's good enough for FSD. That's like me creating a hotdog or not app with only 10% accuracy, and saying my app recognizes hotdogs.
Almost all fsd related discussions have a major logical flaw in that some people here think that being able to "do" or "detect" something means it's good enough for FSD. That's like me creating a hotdog or not app with only 10% accuracy, and saying my app recognizes hotdogs.
powertoold
Active Member
Do you have a paper or article on this? Also, if it's not obvious to you, the sign has NOT been read! You only see the "50", you assumed the rest.
Once again, lidar currently doesn't have the resolution to recognize or read road signs for FSD. You know why? Because of something called vertical resolution at a distance. It seems you like ignoring the obvious limitation in fsd applications.
Last edited:
Given that lidar is moving to beam steering rather than mechanical turntables I would expect the lidar scan can be directed to regions of interest and increase the scanning resolution at that region. Much like the human eye looks directly at objects that it has more interest in seeing. This would enable a targeted approach to the scan picture.
Lumotive Beam-Steering Solid-State Lidar Eliminates Moving Parts
Actually detection and recognition are the same thing. Lidar can detect and classify moving objects, stationary objects, lane lines, lane markings, road markings, curbs, cross walk, traffic lights, etc at accuracy rates at or above that of cameras.
Detecting and classifying vehicles, pedestrians, lane lines, lane markings, road markings, etc are one of the key points of lidar companies that also provide their own software. For example Innoviz..
"The InnovizOne will also support new features such as simultaneous localization and mapping (SLAM) for autonomous vehicle navigation, object tracking, lane detection, and object detection and classification. Innoviz develops.. software to analyze the point clouds generated by its LIDARs. “This software is crucial for delivering superior object detection and classification, detection of drivable area, detection of lane markings and other features that help self driving vehicles understand the scene around them,” Keilaf said.
Yet again you have been proven wrong by independent verifiable evidence and you response with absolute gibberish. I also love how reading road signs is what you latched on to when you said that lidar CAN'T do much of what's on that picture.
Infact this is exactly what mobileye is doing. They detect and classify everything using lidars only no cameras.
But according to @powertoold they have 10% accuracy in their lidar-only system.
The same way a Lidar based perception system can accurately detect and classify vehicles and peds as evident by all SDC companies. Its the same way they can detect and classify all the other things they see in the gifs I posted. The exact same way. If you "disagree", then you have no clue how neural networks work.
powertoold
Active Member
No, plain wrong, particularly in the field of NNs and sensor processing.
Your particular weakness, imo, is that you gather all these marketing information and media and take them at face value. Just because some company can provide a great example of their tech working in some ideal conditions to help sell a product doesn't mean it's good enough for FSD.
A camera is not reading the sign either. Both vision and lidar need software to do that.
powertoold
Active Member
I agree, but in bladerskb's example, you can't even read the sign, as a human. You have to assume it's a speed limit sign. Also, as I said before, lidars don't have enough resolution for FSD sign / road reading because their resolution is very poor at distance.
Lidar can oversample & average, likely it can increase resolution in key areas. Though I'm entirely guessing, but I see no reason to expect it couldn't.
Vision is fixed at 1280x960 at the moment in the Tesla cameras. Likely they can do image stacking also if needed.
powertoold
Active Member
Perhaps, but again, show us the evidence of a current implementation of this.
And here's detection vs recognition for bladerskb:
Object detection versus object recognition
It's kind of a silly semantic difference though, but a clear example is the difference between detecting a road sign vs recognizing what's on the sign (like text). So you have to be careful when some company says they can detect something rather than recognizing it or read it, etc. Another example would be face detection vs recognition.
Last edited:
Its exactly the same.
A neural network when given an image sees 11111 and 00000
Actually they see 0-255
Lane detection is it pointing out which pixels are a lane.
This is the same way with lidar point clouds when fed to a neural network.
Distinction are made in lidar because lidar actually returns points which make up shapes and you can calculate distance and velocity, meaning you can detect objects without the use of a neural network.
This is completely beside the point.
YOU WERE WRONG.
YET AGAIN. YOU REFUSE TO ACCEPT BALANT FACTS.
You went from "an academic exercise to "marketing information and media" and "tech working in some ideal conditions to help sell a product".
Exactly like a typical Ignorant Tesla fan as i expected.
powertoold
Active Member
Not sure what I am wrong about. Best lidar has 250 vertical lines of resolution. This isn't good enough to read road signs or markings at distances required for FSD based on lidar alone without HD maps.
As for your other points on lidar recognizing things, yes, lidar is used to create HD maps, so it's helpful that lidar can assist in detecting things at close distance so that the HD map can be created by the human editor. This has nothing to do with your point that lidar can recognize road markings and signs as well as cameras. It's not even close.
As for your other points on lidar recognizing things, yes, lidar is used to create HD maps, so it's helpful that lidar can assist in detecting things at close distance so that the HD map can be created by the human editor. This has nothing to do with your point that lidar can recognize road markings and signs as well as cameras. It's not even close.
No there isn't. This further shows your ignorance. The term recognition has been deprecated.
The reason its ben deprecated is quite obvious because of its similar word meaning to detection. It is now used in the same vein as detection.
The term used now is Detection and Classification and have been utilized universally for a very long time (years) both in academia and in real world applications.
Last edited:
Similar threads
- Article
