I'm not worked in the AV industry yet so maybe someone in the space can clarify?:
I have a hard time imagining Teslas all sending their entire day's Autopilot footage back to the servers over WiFi during the night. Uploading camera/video data can take quite a bit of bandwidth - at least my household has the largest residential internet plan already and we still get charged overages some months without a Tesla hogging all of it...
e.g. (8 cams) * (1280*964 res) * (60fps) * (16bpp/8) * (1/16 compression ratio) = 74MB per second of footage 13 seconds of footage is 1GB per day => 30GB per month (!!!) And 13 seconds seems enough for disengagements only
Is the compression ratio much higher than 16:1 and if so does this not mess with the ML algorithms due to training on video with lossy artifacts vs. acting on raw frame input in real time?
My point is when people mention that Tesla has the training data of all the millions of miles driven on Autopilot and train on it - "??"s start appearing over my head...
Can someone in the know kindly correct my mental model? Have always been curious on this...
When you’re talking about computer vision data for AVs you need to be talking in TB and PB not GB. The idea that Tesla is offloading high fidelity data at scale is a joke.
I mean idk who you chose to believe or not, but they literally go over this in their presentation. They explain that obviously you don't want to keep all of the driving data because it's just too much, so they have a number of ways to get the data they want. If you don't believe them then that's your problem. But to just say they have no infrastructure to do so is pure speculation on your part, and is disputed by Tesla. And there's no point having a discussion with you because you just choose to ignore any claims they make.
LinkedIn, you really trust Tesla's claims? And also there are people with deeper knowledge of AV (waymo, cruise etc) disagreeing with the whole data thing. Tesla isn't uploading TBs and PBs per day. No one is using only LIDAR. There's a reason every company leading the race is using lidar. They know a thing or two. Give them some credit.
The videos are edited heavily. Drive around bay area and check out the waymo and cruise vehicles. These are the peak of the industry and every hundred or so metres will hit the brakes to a full standstill.
If Intel MobileEye is in there, Nvidia should be it there as well. Its got HW and SW stacks doing exactly the same thing as it. It‘s won several hardware & software deals with major car or truck companies on self driving. Check out the automotive part of its business.
comments
Tier 2: Zoox, ArgoAI, Aurora, Nuro, Uber AGT, Apple SPG and Intel Mobileye.
Tier 3: Lyft Level 5, Optimus Ride, Torc Robotics, Starsky Robotics, Comma.ai, Pronto.ai, Plus.ai, TuSimple, Pony.ai, WeRide.ai, Baidu, nuTonomy, Yandex, Drive.ai, Legacy Car Manufacturers and OEMs etc.
I have a hard time imagining Teslas all sending their entire day's Autopilot footage back to the servers over WiFi during the night.
Uploading camera/video data can take quite a bit of bandwidth - at least my household has the largest residential internet plan already and we still get charged overages some months without a Tesla hogging all of it...
e.g. (8 cams) * (1280*964 res) * (60fps) * (16bpp/8) * (1/16 compression ratio) = 74MB per second of footage
13 seconds of footage is 1GB per day => 30GB per month (!!!)
And 13 seconds seems enough for disengagements only
Is the compression ratio much higher than 16:1 and if so does this not mess with the ML algorithms due to training on video with lossy artifacts vs. acting on raw frame input in real time?
My point is when people mention that Tesla has the training data of all the millions of miles driven on Autopilot and train on it - "??"s start appearing over my head...
Can someone in the know kindly correct my mental model? Have always been curious on this...
But that’s simply moot if they cannot produce the results they claim is possible with what they have.
If they already had what they need to produce results, then they would just show the results instead of just talking.
Tesla presentation was just fundraising..
I think they will be the dominant force in self driving tech.
That helps collect a lot of data and they do have some really smart engineers I’ve heard.
https://www.daimler.com/innovation/case/autonomous/pilot-city-san-jose.html