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
Start a Discussionhttps://teslamotorsclub.com/tmc/tags/

HW2.5 capabilities

Discussion in 'Autonomous Vehicles' started by verygreen, Aug 1, 2017.

  1. AnxietyRanger

    AnxietyRanger Well-Known Member

    Joined:
    Aug 22, 2014
    Messages:
    7,989
    Location:
    EU
    I'm not quite as pessimistic, indeed I think AP2 sees better rearward in such a scenario because it has a far rear fisheye. (It has no nose fisheye so two vans/wagons next to a Model S could block it's front view to the sides.)

    In the rear the biggest vulnerability are the elements. In good weather I think it can probably see enough in faster moving traffic too...

    Still, corner radars in all four corners and a nose fisheye would certainly have helped.
     
  2. selfbp

    selfbp Member

    Joined:
    Aug 4, 2017
    Messages:
    90
    Location:
    East, TN
    New Math will have the ability to solve itself and use pigmentless ink. Pending Regulatory Approval.

    Available for purchase now.
     
  3. _jal_

    _jal_ Member

    Joined:
    Dec 6, 2016
    Messages:
    301
    Location:
    Chicago
    Lidar, radar, and camera pros and cons aside, isn't it pretty clear that the idea was that cameras w/ radar would be processed by a NN? Elon has put a ton of time and resources into deep learning and then gone and included a pretty massive NN processor in every car his company makes. You don't need that kind of hardware to find street signs or even lane lines. Those projects are senior design projects now that you can run on a Raspberry PI.

    If you spend some time and really read the Nvidia work on this, they actually spend a good deal of effort trying to figure out how their trained NN is doing what it is doing in the example project. It just goes to show that there is no algorithmic programming going into it that requires object detection or recognition or classification.

    I wonder what happened / is happening on that front. I thought it was worth a shot given the miles of video Tesla would have and the capital to actually undertake the construction of a training platform up to the task. The Nvidia DGX-1 machines list for $125k a crack, so you can quickly rack up a pretty hefty tab just on hardware.

    I have to imagine they have something up their sleeve to cause them to seek more processing power b/c they certainly can't possible be even close to making out AP2.0 processing power
     
    • Helpful x 1
  4. LowlyOilBurner

    Joined:
    Sep 10, 2017
    Messages:
    14
    Location:
    Buffalo NY
    What gpu is Tesla using now? Drive PX2?
     
  5. Bladerskb

    Bladerskb Like how many times do i have to be right?

    Joined:
    Oct 24, 2016
    Messages:
    721
    Location:
    Michigan
    • Informative x 1
  6. Snuffysasa

    Snuffysasa Member

    Joined:
    Feb 24, 2017
    Messages:
    135
    Location:
    Minneapolis
    Tesla is using the Parker SOC plus an additional discrete pascal GPU. Not sure if the discrete GPU has a name
     
    • Informative x 1
  7. Snuffysasa

    Snuffysasa Member

    Joined:
    Feb 24, 2017
    Messages:
    135
    Location:
    Minneapolis
    This is very interesting thank you for sharing this.
     
  8. stopcrazypp

    stopcrazypp Well-Known Member

    Joined:
    Dec 8, 2007
    Messages:
    9,282
    #508 stopcrazypp, Sep 16, 2017
    Last edited: Sep 16, 2017
    Off topic here, but like @jimmy_d mentions the industry haven't even taken a move to some near-infrared illuminators like security cameras do.
    HW2.5 capabilities

    If night imaging was inadequate, there's a lot more that can be done even with traditional camera technology.

    The comparison uses a IMX174, which is an excellent chip based on native specs (73 dB dynamic range is excellent), but it's not a sensor that comes with native HDR mode (like the Aptina sensor in AP2 with 115+ dB dynamic range with 20-bit output). A properly processed HDR image would do a lot better.
     
  9. lunitiks

    lunitiks ˭ ˭ ʽʽʽʽʽʽʽʽʽ ʭ ʼʼʼʼʼʼʼʼʼ ˭ ˭

    Joined:
    Nov 19, 2016
    Messages:
    1,706
    Location:
    Prawn Island, VC
    • Informative x 3
    • Like x 3
  10. kvandivo

    kvandivo Member

    Joined:
    Sep 11, 2017
    Messages:
    54
    Location:
    Downstate Illinois
    If they have those internships showing up next year I guess that tells us that Tesla doesn't plan to have autopilot solved before Spring 2018.
     
    • Funny x 2
    • Like x 1
  11. Bladerskb

    Bladerskb Like how many times do i have to be right?

    Joined:
    Oct 24, 2016
    Messages:
    721
    Location:
    Michigan
    Funny i actually qualify and can easily get a number of these jobs if i wanted.
     
    • Funny x 3
  12. Ski_

    Ski_ Enthusiast

    Joined:
    Aug 16, 2017
    Messages:
    74
    Location:
    Bay Area
    #512 Ski_, Sep 18, 2017
    Last edited: Sep 18, 2017
    ^^^ Heck, so do I! ;)

    (Jokes aside-- here is the educational requirement for an internship: “Ph.D., Masters and/or Bachelors with significant experience”)
     
  13. verygreen

    verygreen Curious member

    Joined:
    Jan 16, 2017
    Messages:
    1,164
    Location:
    TN
    Nice! We need an internal source, so please do!
     
    • Funny x 2
    • Love x 1
  14. jimmy_d

    jimmy_d Member

    Joined:
    Jan 26, 2016
    Messages:
    188
    Location:
    San Francisco, CA
    I didn't want to get too far off topic so I didn't bring up the FIR camera stuff, but it bears mentioning in discussion of sensor capabilities. FIR cameras have gotten pretty inexpensive (practically free compared to lidar, for instance) and they are a highly complementary to conventional cameras and to radar, providing a lot of data that you can't get in the visible spectrum and having largely independent weaknesses. This also goes for multispectral cameras and broad spectrum cameras which, while not cheap because of their rarity today, would be pretty inexpensive to deploy if anybody bothered to do it in volume. Neural networks operating at the frame level can perform sensor fusion on these various types of cameras very easily so the software development is trivial compared to fusing, say, radar and cameras or lidar and cameras. And yet I don't see anybody doing it. My takeaway from this is that the teams working in this area see so much improvement yet to be had just with visible range cameras that nobody sees any upside right now in augmenting the camera hardware in a significant way.

    Or maybe I'm completely missing something important. That happens too.
     
    • Helpful x 2
    • Informative x 1
  15. zmarty

    zmarty Member

    Joined:
    Jun 25, 2017
    Messages:
    71
    Location:
    Kirkland, WA
    Great find! It seems they are using the Unreal 4 engine to simulate autonomous driving.

    Very similar to what was discussed in episode 71 (and to some degree 70 and 72) of The Tesla Show.
     
    • Helpful x 1
    • Informative x 1
  16. jimmy_d

    jimmy_d Member

    Joined:
    Jan 26, 2016
    Messages:
    188
    Location:
    San Francisco, CA
    High res FIR cameras seem like a good fit to the self driving car application. That said the company seems like they are looking to assemble an aqui-hire team. No meaningful product info, product images are clearly renders, an exaggerated demo, and a press kit that consists mainly of management bios for a bunch of guys with almost no experience in building FIR cameras or automotive anything.

    The side-by-side video demo is really sandbagging the visible spectrum camera. Notice that they have it driving straight into the headlights of a parked car and are allowing camera exposure algorithm to dial the sensitivity way down. No real world system would do this. I notice they end the comparison as soon as the parked car goes out of the FOV.

    But just considering the FIR video by itself shows very nicely how FIR can provide great utility in low light situations. It seems like a great idea. I wonder why we don't hear more about teams using this tech.
     
  17. lunitiks

    lunitiks ˭ ˭ ʽʽʽʽʽʽʽʽʽ ʭ ʼʼʼʼʼʼʼʼʼ ˭ ˭

    Joined:
    Nov 19, 2016
    Messages:
    1,706
    Location:
    Prawn Island, VC
  18. lunitiks

    lunitiks ˭ ˭ ʽʽʽʽʽʽʽʽʽ ʭ ʼʼʼʼʼʼʼʼʼ ˭ ˭

    Joined:
    Nov 19, 2016
    Messages:
    1,706
    Location:
    Prawn Island, VC
    Any news on the code front @verygreen? Rain sensing wipers, maps, radar, camera snapshots, etc.?
     
    • Helpful x 2
  19. BigD0g

    BigD0g Member

    Joined:
    Jan 12, 2017
    Messages:
    773
    Location:
    Somewhere
    Hahahah, you said rain sensing wipers!!!! :rolleyes::rolleyes::rolleyes::rolleyes::rolleyes::rolleyes::rolleyes:
     
    • Funny x 2
  20. verygreen

    verygreen Curious member

    Joined:
    Jan 16, 2017
    Messages:
    1,164
    Location:
    TN
    It looks like the NN for the wide angle cam is in testing (not deployed to the cars yet), possibly it has to do rain sensing wipers.
    Other than that not many news to report.
     
    • Informative x 14
    • Like x 1

Share This Page