Like @Stoneymonster, I also have an AP1 and an AP2 car in the garage. I actually find, that in my local areas, AP2 is now superior at lane detection in most all adverse conditions.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
The Tesla / Mobileye story
- Thread starter lunitiks
- Start date
Like @Stoneymonster, I also have an AP1 and an AP2 car in the garage. I actually find, that in my local areas, AP2 is now superior at lane detection in most all adverse conditions.
I was thinking of next gen AP2+Mobileye, followed by AP2 only. But I could be way off.
Quite likely that Tesla knew deep learning techniques were the future, and they wanted to do it themselves. Mobileye autopilot was a band-aid. However, that band-aid was ripped off sooner than Tesla anticipated.
Yep, plausible.
Also, AP "1.5" was likely planned as second-gen band-aid was for some reason - a story not told - it didn't make it. Delayed too long and then AP2 came?
And yes @mongo, if that's what you meant I agree AP2 was originally also slated to have EyeQ3... until it would no longer be needed. See the empty spot on the AP2 board...
googleiscoul
3 months maybe, 6 months definitely......
I'm pretty sure AP2 can distinguish between bicyclists, and motorcycles it just seems they haven't started rendering them on the dashboard just yet. (skip to 4:57 to see how the PX2 sees the world)
We have no reason to believe AP2/EAP runs that codebase yet, though.
googleiscoul
3 months maybe, 6 months definitely......
True I have heard that Tesla runs a very highly customized version of that software or they use their own in house software and it is very possible that it might not behave in the same ways as shown in that short demo.
I thought AP2 currently only showed reindeer that are in your personal strip of ice. AP1 shows deer on either side as well...
That's really impressive, because AP1 has gotten much, much better over the course of the last year. If AP2 is matching the current state of AP1, it's far better than the AP1 I had when I got the car with 7.1 on it.
Nothing to add here but just a thank you. That video a few posts up was fascinating. Thank you to the OP that created this thread and to all that contributed to it.
Thank you for this post.... The over idealization of mobileye and every other autonomous car program not yet meaningfully deployed is going to drive me insane.... especially if mobileye and cadillac supercruise fans have to come on a tesla forum to declare tesla's method overly simplistic or flawed.... .... It reminds me of apple hater snobbery. I get it, there are other more capable devices, but until someone actually explains how unbelievable an achievement it is that apple has achieved the unthinkable in finding a way to get tech averse grandparents to FaceTime with ease... I just can't take haters seriously.
Tesla has said over and over again, that they are trying to deploy self-driving to the masses for little money with less processing power. Hotz is the only other person that speaks about that critical point. @lunitiks, save me
are you people really this naive? Hotz is as clueless as they come.
He keeps saying that Tesla is using a deep learning approach different from everyone. But Tesla isn't doing any more of a deep learning approach than anyone else. they are using as much machine learning as everyone because every one uses deep learning for sensing just like Tesla.
Hotz is clueless.
He tweeted this months ago
comma ai on Twitter
This is a mobileye 2 year old talk that i have posted here a hundred times and he is now finally discovering it and responding to it. That shows you how clueless Hotz is. I have talked to Hotz several times on reddit and you can easily tell that he knows nothing about the autonomous industry.
If you actually watched Hotz you would know that everything he has publicly criticized Mobileye for he has backpedaled.
First he said HD Map was useless and pointless, now he's building HD maps and saying its necessary..
Second he said there was no need for specific domain detection models, now he is building semantic free space detection and bounding boxes, etc.
comma ai on Twitter
He even announced today that he will be doing traffic light and stop sign detection and stopping and he will be using high precision gps and a user maintained map of traffic lights/stop signs.
Everything he was against before and saying to gain media attention. Now he's realizing he's been wrong the entire time.
Yet he still praises Tesla for PR reasons because of the huge media attention he gets when he does it. He keeps claiming he and Tesla are the only one going the machine learning when he's clearly not.
Wake up people.
Well Wow @Bladerskb, you answered one of my questions. Hotz walked it back. I'm surprised you don't like Hotz though.... you have the same style....you know... sweeping and scathing judgments, over simplifications and pontifications. It's entertaining as hell, but just so damn ironic that the you two both remind me of Macho man.. at least you are funny. Hotz's old interviews make me want to puke as he seems like a character on silicon valley. I think Hotz's drastic characterizations on mobileye had everything to do with feeling like musk chose them over him or something... anyway, glad you came out of your Kremlin office to post this.
@buttershrimp not sure what to save you from, you're doing an excellent job here.
Btw in your next vid, I request O.T. Genasis - CoCo!
I thought AP2 currently only showed reindeer that are in your personal strip of ice. AP1 shows deer on either side as well...
You got me there.
I'm slipping.
Another thought:
One of the real benefits of deep learning is the ability to learn low and high level features automatically from the data (as opposed to hand-crafted). Of course this is a complicated process for self-driving requiring lots of data that may be sparse, need labeling, etc...
My guess is for EAP 2.0, instead of learning against some output driving cost function, Tesla was going to initially train deep nets against the output stream from Mobileye, and in essence, recreate those complicated hand-crafted features into their own NN architectures and weights, that Tesla owns. Or, they would use the Mobileye stream to label their camera data, so that they could, for example, throw out 1000s of hours of repetitive scenes but find and keep all the edge cases, with less manual input.
Either way, this would have significantly accelerated the learning and advancement of AP 2.0. It's not surprising that Mobileye would refuse to participate in this.
One of the real benefits of deep learning is the ability to learn low and high level features automatically from the data (as opposed to hand-crafted). Of course this is a complicated process for self-driving requiring lots of data that may be sparse, need labeling, etc...
My guess is for EAP 2.0, instead of learning against some output driving cost function, Tesla was going to initially train deep nets against the output stream from Mobileye, and in essence, recreate those complicated hand-crafted features into their own NN architectures and weights, that Tesla owns. Or, they would use the Mobileye stream to label their camera data, so that they could, for example, throw out 1000s of hours of repetitive scenes but find and keep all the edge cases, with less manual input.
Either way, this would have significantly accelerated the learning and advancement of AP 2.0. It's not surprising that Mobileye would refuse to participate in this.
jimmy_d
Deep Learning Dork
That's a clever approach; doing transfer learning from an old system to a new one. The training might take a while if you do it with hardware - it's usually done algorithm-to-algorithm inside the same fast machine to save calendar time. But it would probably work and it probably would have been a quick way to duplicate the AP1 functionality in the AP2 hardware.
But it's very likely that the IP agreement between ME and Tesla would have prevented this. Reverse engineering is almost always prohibited in these kinds of contracts and you can be sure that Tesla using transfer learning to duplicate the ME functions in the presence of a no-reverse clause would bring a lawsuit. On top of that it seems likely that Tesla's ginormous self confidence would lead them to believe that they can outdo ME even without a head start.
Still, it's a great observation.
Similar threads
