Terminator857
Active Member
You have limited resources. You have more projects to work on than can be done. Projects must be cut. LIDAR out, radar out, etc...
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Lidar vs Camera revisited
- Thread starter heltok
- Start date
EVNow
Well-Known Member
I don't think its a disadvantage - except when Tesla started it was. Ugly Lidar couldn't be added to consumer cars.
You would still have sensor integration issues ... but thats another story.
In spite of what Elon says - I believe Tesla would have gone with a richer suit of sensors if they were starting out today.
Daniel in SD
(supervised)
Yes, but OP seems to be saying you're still going to generate a point cloud with cameras. If that problem had been solved then no on would use LIDAR at all. It's an extremely difficult problem and something that Tesla is devoting resources to.
linux-works
Active Member
thing is, if tesla kept their once-stellar rep, they'd have zero problems finding people to work there.
they'd have to pay well. they'd have to stop the texas BS (turns a lot of progressives off, and we are the key demographic for your technology, DUH!), they'd have to admit when they were wrong, they'd have to start taking customer support/care seriously and they'd have to fire that madman (or just get rid of him) and find a sane leader who has realistic goals and proper use of corp funds.
none of that is going to happen any time soon.
tesla has peaked and while its sad, its a cycle and its how the world works.
what IS sad is that this was entirely preventable. if they only had competant management and not an egomaniac.
Daniel in SD
(supervised)
I understand why they didn't add LIDAR to consumer cars. Camera only is fine for driver assist and safety features. I bet most cars will dump radar and go camera only.
EVNow
Well-Known Member
Thats good information. It also makes you an employee of potential Tesla competitor. You should note that in your signature.
Depends on your idea of what the car error rate can be. If you think about it - there is so much chaos on the road that there should be a lot more accidents. And then you go and see real chaos on the roads of Cairo or Chennai. Then you wonder how come there aren't accidents every second in every road.
Those in the "field" also thought you can't make profitable EVs or make people actually want to buy EVs. 'nuff said.
You are just saying the same thing we have all been talking about for years here. Nothing new.
ps : working on projects that you are sure will fail, is one way to make sure the project fails.
Terminator857
Active Member
Please clarify. I suspect you are talking about Level 2 cars. Not level 3 and or level 4.
Terminator857
Active Member
I don't understand. ML feature engineering is not done this way.
Feature engineering is done by their engineers.
You'll have to explain what you mean by feature engineering. Are you talking about labeling?
Last edited:
Daniel in SD
(supervised)
Of course. I don't expect Level 3 or 4 consumer cars in the US anytime soon.
Terminator857
Active Member
Agree, but the are plenty that are advertising they will upgrade to l3 and l4 when the software is ready. Chinese cars, german luxury brands, volvo etc...
Example:
2022 Volvo XC90 Will Get Autonomous Highway Pilot Hardware with Lidar
The next-gen XC90 will be the first car to be equipped with the necessary Luminar lidar hardware, but the software to make the car autonomous won't be ready till later.
www.caranddriver.com
Last edited:
Daniel in SD
(supervised)
I'll believe it when it's actually released. I think we'll see L3 highway systems that use LIDAR in other countries.Agree, but the are plenty with that are advertising they will upgrade to l3 and l4 when the software is ready. Chinese cars, german luxury brands, volvo etc...
Example:
2022 Volvo XC90 Will Get Autonomous Highway Pilot Hardware with Lidar
The next-gen XC90 will be the first car to be equipped with the necessary Luminar lidar hardware, but the software to make the car autonomous won't be ready till later.
Feature engineering refers to the process of using domain knowledge to select and transform the most relevant variables from raw data when creating a predictive model using machine learning or statistical modeling. The goal of feature engineering and selection is to improve the performance of machine learning (ML) algorithms.
What is Feature Engineering? Definition and FAQs | OmniSci
Learn the definition of Feature Engineering, and get answers to FAQs regarding: Steps in feature engineering and how it is used for machine learning models.
www.omnisci.com
Examples of this could be to transform the raw sensor data into a voxel grid, do clustering, find nearest neighbors of point and generate a graph of these, NARF etc.
Last edited:
Terminator857
Active Member
Yeah, feature engineering is done by engineers / scientists, not in the automated pipeline as your original post suggests.Feature engineering refers to the process of using domain knowledge to select and transform the most relevant variables from raw data when creating a predictive model using machine learning or statistical modeling. The goal of feature engineering and selection is to improve the performance of machine learning (ML) algorithms.
What is Feature Engineering? Definition and FAQs | OmniSci
Learn the definition of Feature Engineering, and get answers to FAQs regarding: Steps in feature engineering and how it is used for machine learning models.
Representation: Feature Engineering | Machine Learning | Google for Developers
developers.google.com
Last edited:
It is developed(done) by engineers and then run in systems either online or offline in an automated pipeline.Yeah, feature engineering is done by engineers / scientists, not in the automated pipeline as your original post suggests.
Representation: Feature Engineering | Machine Learning | Google for Developers
Last edited:
Terminator857
Active Member
"run in systems" means simple math. It is not feature engineering at that point. When Karpathy says training for those features, he means optimization for the goals of that neural network, which is based on features. Features can be simple features or engineered features. An engineered feature can be to lower the jerk which is the third derivative of ( distance / time ).
Last edited:
Terminator857
Active Member
Current project is more software engineering than ML, but expected to more towards ML next year. I've done some light assistant teaching of ML.
The thing that bugs me about relying on Vidar or Pseudo Lidar is that there is just a little too much magic involved. When you are heading directly at a unresolvable object - like a solid white wall - the technology is useless. Lidar or Radar would still say "Object Ahead".
If the situation demands life-critical sensing, I would say use the best sensor, don't Pseudo it with fancy computerized assumption software.
If the situation demands life-critical sensing, I would say use the best sensor, don't Pseudo it with fancy computerized assumption software.
I didn’t say it was $40. I said nio ET5’s cost $40k.
High resolution lidar in mass production today cost anywhere between $150 to $500
Power? lol the average power consumption of lidar are 10 watts.
Compute? These cars have anywhere from hundreds of TOPs to thousands depending on the company. They are not compute constraint, secondly processing lidar input requires way less compute than cameras.
All I see is made up points to say lidar is worthless. Just like Elon. Note if anyone tells you something is worthless that everyone else uses and you couldn’t until very recently...They are pushing an agenda.
Lastly it’s good to point out that No one is of the opinion that cameras wouldn’t eventually be enough for safe FSD. But that it will take till around 2030 as ML improves. This Is another argument derived from exactly the opposite of what ppl actually believe.
Similar threads
- Article
