TMC is an independent, primarily volunteer organization that relies on ad revenue to cover its operating costs. Please consider whitelisting TMC on your ad blocker and becoming a Supporting Member. For more info: Support TMC

Seeing the world in autopilot, part deux

Discussion in 'Autopilot & Autonomous/FSD' started by verygreen, Sep 25, 2018.

  1. kdday

    kdday Member

    Joined:
    Oct 29, 2016
    Messages:
    743
    Location:
    AZ
    Well I know @BigD0g purchasd FSD and he's complaining his pillar cameras aren't calibrated. I'm wondering if it's a hw2 vs hw2.5 thing....
     
  2. croman

    croman Active Member

    Joined:
    Nov 21, 2016
    Messages:
    3,799
    Location:
    Chicago, IL
    Exactly. It's the dashcam thing for 2.5 that calibrates the cameras. Are there active NNs for all 8 now? If so then maybe there is credence to FSD feature development.
     
  3. croman

    croman Active Member

    Joined:
    Nov 21, 2016
    Messages:
    3,799
    Location:
    Chicago, IL
  4. AnxietyRanger

    AnxietyRanger Well-Known Member

    Joined:
    Aug 22, 2014
    Messages:
    9,451
    Location:
    EU
    #124 AnxietyRanger, Sep 27, 2018
    Last edited: Sep 27, 2018
    Tesla did put traffic light and stop sign imagery into the firmware years ago, and is yet to launch any such feature. While an interesting find, I'm not sure it directly correlates into any kind of information on Tesla's object-direction recognition prowess? I agree they could be seeing the direction, but I'm just saying we hardly have much evidence of it yet?

    8.1 certainly does not reliably see the direction of an object, which would correlate with the 2D bounding box finding, much better than with a "3D box in secret" thesis. I'm not sure where this thesis comes from for the leaked developer version of this thread, except perhaps wishful thinking at this stage? V9 of course is still an unknown quantity, but it doesn't seem impossible it might still extrapolate object direction from simply direction of travel or lane markings, instead of visual recognition?

    There is the theory that AP2 GPU/CPU setup - prior to any possible "AP3" GPU/CPU upgrade - are not powerful enough to run eight cameras (and not the dash cam either). If @BigD0g is right, that certainly could explain it. After all, what V9 seems to be offering is the first version of the lane-changing EAP, which Tesla originally promised to do with just four cameras...

    That said, I agree there certainly seems to be evidence of it using more than four - so maybe AP2 is powerful enough for the types of NNs they are running at this stage. Just pondering. Originally eight-camera fare was an FSD feature, EAP was advertised as four-camera.

    I do think Tesla is pushing the industry - in a new direction. I believe you too have made the point that the rest of the industry has been looking beyond aggressive Level 2 implementations, but have been forced to re-think this approach due to Tesla taking Level 2 far beyond where they had planned - because, as you say, their eyes and development projects were always aiming for Level 3 and beyond (i.e. car responsible, not driver). You are of course right Audi has shown very impressive self-driving cars already years and years ago, but their conservative launch schedule is... conservative.

    As I mentioned in my previous lenghty post, I do NOT think Tesla's advantage is their vision progress - the status of that is a separate question and in all likelihood the likes of Waymo and MobilEye are significantly (and that's putting it mildly) more advanced and stable in that regard. This is likely why Supercruise is so stable, for example. Tesla very likely has to factor in a much higher rate of fails in their vision at this time, than either Waymo or people using MobilEye products. They also don't have so many radars (or lidar) to confirm their visual recognitions against, compared to much of the competition.

    Tesla's advantage - and where they are pushing the industry - is how aggressively they have moved even with those limitations - first lacking in sensors in AP1 and now with arguably an immature vision engine in AP2+ and again lacking in non-vision sensors. They are also aggressive in the number of cameras shipped for ADAS.

    I think it would be a service to this discussion if we could keep three things separate:

    1) Tesla's vision software status (IMO not mature yet compared to the competition from MobilEye and Waymo, but this may be offset by other factors - see point 3)

    2) Tesla's/industry's FSD/SDC status (Waymo clearly leading the pack - Tesla's ability to scale to Level 3+/car-responsible quite unknown given their driver-responsible approach and "beta status" in point 1)

    3) Tesla's/industry's Level 2 status (Tesla pushing consumer-available ADAS through larger camera suite and aggressive software development, arguably has a leading position - the "what consumers can actually buy" difference)

    In essence, all these things - just hypotheticals - can be true at the same time, because those three things are separate:

    - MobilEye has the best automotive vision solution by far
    - Waymo has the best self-driving car by far
    - Tesla has the best ADAS consumer money can buy by far

    Keep up the good work, @verygreen and @DamianXVI! And keep your interesting points coming @Bladerskb too. Thank you.
     
    • Like x 4
  5. BigD0g

    BigD0g Active Member

    Joined:
    Jan 12, 2017
    Messages:
    1,898
    Location:
    Somewhere
    No complaints, just an observation :rolleyes:
     
    • Like x 2
  6. Bladerskb

    Bladerskb Senior Software Engineer

    Joined:
    Oct 24, 2016
    Messages:
    1,676
    Location:
    Michigan
    Really? I didn't know US release date for Spring 2019 is 2020..which by the way it clearly says in the article. That must be some bubble you are in.

    And as you couldn't tell (not surprising)

    That 2019 bmw has trifocal focal camera and is just one of the two cars coming out in 2019 that will use eyeq4 and trifocal (I surmise the potential L3 highway nissan car in japan is the other one)

    There will be total 4 cars in 2018 with eyeq4. 12 more in 2019 with eyeq4. 2 of the 12 will use trifocal cameras. All has paid for full feature bundle chip.
     
  7. J1mbo

    J1mbo Active Member

    Joined:
    Aug 20, 2013
    Messages:
    1,420
    Location:
    UK
    Trifocal is so 2017
     
    • Funny x 1
  8. croman

    croman Active Member

    Joined:
    Nov 21, 2016
    Messages:
    3,799
    Location:
    Chicago, IL
    4 cars with eye Q4 by the end of the year. Mmmk.
     
  9. chillaban

    chillaban Active Member

    Joined:
    May 5, 2016
    Messages:
    3,595
    Location:
    Bay Area
    This year maybe next year definitely.
     
    • Funny x 5
  10. strangecosmos

    strangecosmos Non-Member

    Joined:
    May 10, 2017
    Messages:
    1,041
    Location:
    The Prime Material Plane
    Does Mobileye collect sensor data or other driving data that could be useful for training neural networks that do perception (e.g. depth mapping, semantic segmentation, instance segmentation) and prediction? Or to train the reinforcement learning algorithms Mobileye wants to use for path planning and/or control?

    To me, that is the real showstopper. If I found out Mobileye is collecting data from these EyeQ4 cars with trifocal cameras the same way Tesla is (not just for HD maps), then that would change my view of Mobileye considerably.

    My basic thesis is that the development of full autonomy is a function of training data. So is Mobileye is collecting lots of training data from production cars, it would stand to reason that Mobileye has an advantage in speed of development.
     
  11. Snuffysasa

    Snuffysasa Member

    Joined:
    Feb 24, 2017
    Messages:
    271
    Location:
    Minneapolis
    #131 Snuffysasa, Sep 28, 2018
    Last edited: Sep 28, 2018
    It's possible Mobileye could collect training data from production cars, but I doubt they will.

    But also Tesla is hardly collecting training data from their production cars.


    You are vastly over valuing training data. Nor mobileye nor Tesla's development of full autonomy is limited by the amount of unlabeled data. (or even labeled data for that matter)

    Furthermore, there are hundreds or thousands of challenges that go into developing full autonomy, training the NNs is just one.

    And about reinforcement learning, you cannot use fleet data for reinforcement learning, the fact that you even suggest this is a possibility makes it clear you do not understand how it works.


    lol
    Lol, nice thesis. Find me one computer vision / perception engineer at Tesla or Mobileye or Waymo, or any other that says their development is currently labeled by the amount of fleet data they can collect.
     
  12. chillaban

    chillaban Active Member

    Joined:
    May 5, 2016
    Messages:
    3,595
    Location:
    Bay Area
    I think you’re underestimating certain types of training data though. The ability to perform snapshot triggers on customer cars as well as aggregate performance statistics is huge. I’ve been in that line of work before where we have a huge engineering fleet collecting a ton of unimaginably intrusive data but still, some customer with a certain configuration in a certain location can make things happen that take man-months for us to discover.

    Yes the billions of miles number is marketing fluff but that doesn’t mean it’s worthless to have a fleet of customer cars all over the world that can upload data to you at will.
     
    • Informative x 3
  13. pkodali

    pkodali Banned

    Joined:
    Oct 22, 2017
    Messages:
    817
    Location:
    Sacramento
    Looks like this is the view from the Tesla dashcam

     
    • Informative x 4
    • Like x 1
  14. Snuffysasa

    Snuffysasa Member

    Joined:
    Feb 24, 2017
    Messages:
    271
    Location:
    Minneapolis
    I agree there is value to it.

    However, that value is NOT collecting data useful for training data for training NNs.


    Strangethecosmos was talking about mobileye, you can get some value from a fleet, however that is value to the Automakers, not to mobileye
     
  15. strangecosmos

    strangecosmos Non-Member

    Joined:
    May 10, 2017
    Messages:
    1,041
    Location:
    The Prime Material Plane
    #135 strangecosmos, Sep 28, 2018
    Last edited: Sep 28, 2018
    The way Andrej Karpathy tells it, making HW2 Teslas autonomous is mostly a matter of neural networks, and neural networks are mostly a matter of creating data sets. He says neural network architecture is much less of a focus for him than data sets. In the video, Karpathy is specifically talking about deep supervised learning for computer vision neural networks.

    If a company wanted to use production fleet data to train a neural network for path planning, it might be able to use sensor data to capture real world situations that can be re-created in simulation. This would help the simulated cars, pedestrians, bikes, etc. reflect the behaviour of the real entities. They could then use reinforcement learning to train the neural network on path planning in simulation.

    A company could use driver disengagements from ADAS like Autopilot to flag the real world situations where path planning fails and needs additional training in simulation, or in structured tests on private roads.

    What do you think are the main bottlenecks in developing full autonomy?
     
    • Like x 1
  16. verygreen

    verygreen Curious member

    Joined:
    Jan 16, 2017
    Messages:
    2,392
    Location:
    TN
    narrow cam (to hide lack of resolution?, but the tradeoff is you don't see what's on the sides, and it shakes more due to bigger focal distance)
     
    • Helpful x 3
    • Informative x 1
  17. pkodali

    pkodali Banned

    Joined:
    Oct 22, 2017
    Messages:
    817
    Location:
    Sacramento
    Wonder if the will overlay this over the normal front camera one for the next release, allowing for an area in the enter with higher resolution and less resolute outer region
     
  18. strangecosmos

    strangecosmos Non-Member

    Joined:
    May 10, 2017
    Messages:
    1,041
    Location:
    The Prime Material Plane
    • Like x 1
  19. Bladerskb

    Bladerskb Senior Software Engineer

    Joined:
    Oct 24, 2016
    Messages:
    1,676
    Location:
    Michigan
    I disagree in that, what's AP kill counter up to again? 4? 5? and how many accidents hundreds?

    Do you believe that traditional automakers are motivated by that? Every News brake about another death in AP, do you think they go, "Yes we want to be like them!!"

    I think its the opposite, Tesla went from allowing people to go a 1,000 miles without touching the wheel to now going 30 seconds without touching the wheel. That's the biggest backtracking of the century.

    In retrospect, GM pushed the industry by using driver-facing camera and HD maps and allows you to go potentially hours without touching the wheel. Now everyone is trying to follow suit.
     
    • Disagree x 2
  20. bebbiXpress

    bebbiXpress Member

    Joined:
    Dec 12, 2017
    Messages:
    158
    Location:
    Basel
    I respect an objective and constructive criticisms of any system but what you produce is simply FUD. You're a Mobileye fan-boy forgetting that a lot of accidents and the majority of deaths were a result of an inattentive driver and Mobileye's AP 1 system too limited to (reliably) recognize non-moving or cross-moving vehicles or objects.

    There are 3 AP deaths Wikipedia knows of, 2 happened with Mobileye's AP 1! List of self-driving car fatalities - Wikipedia If you can prove there are 1 to 2 more, please provide a reference. Same goes to the hundreds of AP accidents, prove it or the real numbers are in the tens and not hundreds.

    I know, you're now going to blame Tesla for misusing Mobileye's system, blabla...
     
    • Like x 2
    • Love x 2
    • Informative x 1

Share This Page

  • About Us

    Formed in 2006, Tesla Motors Club (TMC) was the first independent online Tesla community. Today it remains the largest and most dynamic community of Tesla enthusiasts. Learn more.
  • Do you value your experience at TMC? Consider becoming a Supporting Member of Tesla Motors Club. As a thank you for your contribution, you'll get nearly no ads in the Community and Groups sections. Additional perks are available depending on the level of contribution. Please visit the Account Upgrades page for more details.


    SUPPORT TMC