Research at NVIDIA: Structured Domain Randomization

[Interestingworld]

1:43 yeah why the bus was no detected??🤔

[comment.highlighted]

01:42 How come the Bus was not detected? 🤔

[Ouyk]

Randomize nearly everything: lane marking paint, potholes, pedestrians, obstructions, signs, vehicle size/shape, bridge size/shape, street and freeway size/type/shape, weather, daylight, traction, speed, loss of sensor information, map information, buildings, bicycles, motorcycles, charge ports, etc. Run your simulations on a supercomputer for a few trillion miles in a state or county full of autonomous and non-autonomous vehicle data from real cars. If there are almost no accidents I'd actually trust a nVidia-equipped autonomous vehicle.

[mb]

1:30 posting that video evidence of you driving on the shoulder eh

[flintstone]

1:40 welcome on the Standspur, which is not made for driving on it

[vicproon]

did you use imagenet-pretrained feature extractor?

[ScrapTechTips]

Well here comes a learning thing we cannot ignore

[merbst]

I hope I am not the only one who sees terrifying philosophy implications from this.

[theendurance]

This is what separates NVidia with AMD. Nvidia puts more effort on real world uses and the future, while AMD is just focused on gaming/workstation

[Navhkrin]

Beautiful, this will change the game by eliminating the need for vast amounts of human labeled data.

[hennadiikasian]

Why you re-inventing vehicle? You can use excellent data models from Unity 3D (which already contains some background for migration ML solutions). Or even UE 3 for the similar tests. Both engines contains sky, projections, mirrors, lights, random worlds generation and best performance.

[zqa]

Looks like I'll be out of a job soon :p

[Mart-E]

Teach it to detect buses and trains too

[TheGUARDIANOFFOR]

LOL many cars are ignored or dected on last second.... THIS IS

[broForseNetwork]

Thu it's not as good as Tesla v9 they can tell the different sides of the car and draw a 3d box around ir