transcribe

GTC 2016: Deep Learning for Cars (part 10)

NVIDIA · 9m · transcribed 23d ago
More from NVIDIA Business
𝕏 Share ▶ YouTube 📥 PDF 🤖 .md

Transcript

0:00 let me talk about one more thing the last thing is deep learning for cars well if it was an amazing year for AI it's also an amazing year for cars there are just so many announcements the fact of the matter is we all know this AI is coming to cars AI is coming to cars when we thought about helping to create self-driving cars in the beginning the basic approach was sense plan act using traditional computer architectures using traditional

0:31 Computing techniques we're going to continue to do that as well we're going to continue to do sense plan act the basic Loops of self-driving cars is know where you are in the world figure you know based on a map also based on your census localized you see what's around you and then you you you act plan sense sense plan act the basic Loop of Robotics so we decided to build a computer dedicated for that very functionality we

0:58 built the world's first AI car computer it's called the Nvidia Drive our basic approach was this we would build an endtoend Computing platform all the way from training the network that would ultimately run in your car training the network is the beginning training your network for your car as you've heard from all of the leaders training your network is the first step of having an autonomous driving car and so we've created a platform that allows you to train on any framework you

1:27 would like to use Cafe cntx cntk that is excuse me tensor flow theano torch you name it training your network and then taking that exact same network running on something that has incredibly high performance but has also very high Energy Efficiency as I mentioned today with gie you could take that Network and run it on a car computer we call that car computer the Nvidia Drive PX it's the world's first deep learning car Computing platform it's scalable we

1:55 believe that AI will be used for car Computing not just for autonomous driving AI will be used to talk to you you would give it commands it'll be looking around in your smart mirrors your car will be able to see all around you and tell you there's a bicycle on your right of course it's going to avoid it anyways but it's going to tell you you could talk you could talk to it you could change temperatures

2:16 it looking at you looking at your gesture inside the car it could tell when you're dozing off it could tell whether you're looking in the direction of the oncoming traffic AI will be used inside the car AI will be used to drive the car AI will be used to keep you out of Harm's Way We Believe AI will be used across the board and car computers we used one architecture from infotainment to Cluster to Adas to

2:41 self-driving all the way to mapping it's completely open platform all the apis are available to everybody now the last time I showed you we have started working on object recognition for our Drive computer the thing that's really amazing is only in just a few months after a great deal of effort our scientist s our researchers working on our network has now achieved the number one score on the kitty Benchmark this is a self-driving Benchmark I'm super proud

3:08 of them for doing this we have the number one accuracy on Kitty the thing that's really great is that number two number three number four number five number six number seven number eight all GPU accelerated working with a large community of researchers that are trying to develop self-driving cars we're open to work with everybody we want to encourage everybody to enable this capability so we can make streets safer so we make driving better and hopefully

3:39 it might even change the way we design our cities well let's take a look let's take a look at what it can do okay so not only is it accurate as I mentioned earlier gie also makes our Network running on our processor super fast and so we're detecting cars in the front on the upper right in the back or no front in the upper left and behind you in in the upper right but also in your side mirrors so the thing that's

4:05 really cool is that you could detect objects all around you using exactly the same network and all of that is running on this little tiny processor the small version the smallest version of Drive PX this is the smallest version of Drive PX 180 frames per second 180 frames per second detecting cars all around you you now we can detect Lanes we can detect detect signs we can detect anything we wanted to train it to do and that's really the

4:36 wonderful thing that's the beautiful thing about a deep neural network whatever you train to do it will do whatever you try to detect it will detect okay so Drive PX well the next thing we like to do so that's an update on perception we like to also make sure that you can map well if you want to figure out exactly how to drive and drive safely you want to have as much information as you possibly can

5:03 this is called HD mapping you could use lar but we also believe that you ought to use l art and camera and everything that you have and the cameras will use photogrametry basically F motion photogrametry to use to do basically what is called structure for motion to create and identify important points in this in the in the surrounding and map reconstruct a 3D World for you okay that mapping is called high high definition mapping and we've created a platform that allows

5:32 us to do that let's take a look at that oh can we well I think I just leaked that out out the door okay excuse the sequence so by by do Andrew Ang and his lab as you guys know are working on a self-driving car and he he was so enthusiastic about it he asked me to to share with this with you guys the thing that's really

6:02 important is this while you're creating a self-driving car you got all kinds of algorithms so you want to deploy onto it you have sensors and cameras all over your car what kind of computer would you need well it turns out they need a GPU accelerated supercomputer and before the existence of dve px2 they had to build their own they had to build their own but they have wonderful computer scientists they have wonderful computer engineers and they built a cluster a

6:26 supercomputing cluster that fits in the trunk of a car now this isn't something that every car company can do and this isn't something that every car company wants to do and so the thing that we did was we took all of that supercomputing horsepower And We Shrunk it into one platform we call drive px2 drive px2 has two te processors on it and it has two Pascal Next Generation gpus that hasn't been announced on it and so if you take a picture of that and

6:56 you zoom in carefully which I know you guys will do you do it to me all all the time let's see hang on a second I took the serial numbers off okay and so this four chip solution whereas you were looking at a one chip solution just now this four chip configuration of the architecture would allow you to basically have a supercomputing cluster inside cars you connect 12 cameras to it connect it to your Camus and you can be doing pH

7:25 photogrametry motion photogrametry as well as col collecting information from lighter fusing it in real time Computing the map and instead of streaming up to your Cloud 12 highdefinition video streams you stream up to your Cloud Point clouds that are highly compressed and that will be accumulated into Cloud over time crowdsourced by all these cars that are driving around registered in account Cloud to create one big high definition map and so this particular

7:56 platform we're announcing today will also be a mapping platform the car computer will be Drive px2 the computer that's in the cloud will be djx1 n to n exactly the same architecture px2 in the car DX gx1 in the cloud let's take a look at let's take a look at real-time mapping so basically What's Happening Here is that this car our car this is using four cameras it's able to detect

8:29 up to 15,000 important points it selects 15,000 important points per frame per second per camera and so altogether the four cameras can collect up to 1.8 million points per second 1.8 million points per second the points are not only in 3D space but it is also in color and so we would accumulate all this information and we would load it up into the sky and that will be then registered and calibrated and maintained using djx1 okay so this is ma mapping we

9:01 created this platform so that every mapping company in the world could have the benefit of mapping cars building these car computers for mapping are really it's a really expensive Endeavor you're basically creating a supercomputer that fits into the trunk of a car but we've now with drive px2 with djx1 have created an endtoend Computing platform that car computers can use mapping companies can use and we have to map the world before all the cars can safely

9:31 travel around the world

© transcribe · For agents Built with care and craft by Gokul Rajaram