Nvidia's 2015 CES press conference | CNET

Nvidia's 2015 CES press conference

The graphics chipmaker will talk up its latest efforts in automotive and mobile.

  • Ben Fox Rubin 1/5/2015 4:48:50 AM
    Important part of this is recognizing a pedestrian, even if only partially seen.
  • James Martin 1/5/2015 4:49:18 AM
  • Ben Fox Rubin 1/5/2015 4:49:23 AM
    The machine can tell "it's likely a human" just by seeing part of a person's image.
  • Ben Fox Rubin 1/5/2015 4:50:08 AM
    Can tell the difference between a school bus or car. The vehicle would become even "more alert" around a school bus, Huang said. Or the vehicle could pull aside when recognizing an ambulance.
  • Shara Tibken 1/5/2015 4:50:14 AM
    Now we're getting some talk about computer vision. Pretty interesting area of tech. Facial recognition, etc, is being worked on by the big tech companies like Google. In the context Nvidia's talking about, it's for detection in cars. You can tell someone partially blocked by something else is still a person whereas a normal computer may not know that.
  • James Martin 1/5/2015 4:50:20 AM
  • Ben Fox Rubin 1/5/2015 4:50:28 AM
    But to do this stuff, need to teach a machine to recognize all this stuff in the world.
  • Ben Fox Rubin 1/5/2015 4:50:50 AM
    Showing video demo of Drive PX after being trained on seeing what's out there.
  • James Martin 1/5/2015 4:51:23 AM
  • Ben Fox Rubin 1/5/2015 4:51:27 AM
    First step, detecting crosswalk signs, which in the video get green squares around them.
  • Ben Fox Rubin 1/5/2015 4:51:52 AM
    Those green boxes show that the system can see these signs.
  • James Martin 1/5/2015 4:52:23 AM
  • Ben Fox Rubin 1/5/2015 4:52:23 AM
    Now going through detecting different speed limit signs, pedestrians.
  • Ben Fox Rubin 1/5/2015 4:52:41 AM
    One of the more difficult is a partially viewed pedestrian.
  • James Martin 1/5/2015 4:52:57 AM
  • Ben Fox Rubin 1/5/2015 4:52:58 AM
    Can also detect street lamps.
  • Ben Fox Rubin 1/5/2015 4:53:16 AM
    By the way, it's not a live video...but the processing is live, Huang said.
  • Ben Fox Rubin 1/5/2015 4:54:11 AM
    Car can tell which lanes can stop or go by reading street lights.
  • James Martin 1/5/2015 4:54:33 AM
  • Ben Fox Rubin 1/5/2015 4:54:57 AM
    Once it sees the lights, it can subcategorize them as red, yellow or green (in some cases seeing by position of lights).
  • Ben Fox Rubin 1/5/2015 4:55:36 AM
    Can more easily see the head-on lights, some are registered as unknowns.
  • Ben Fox Rubin 1/5/2015 4:55:49 AM
    Hard-lighting conditions are especially tough.
  • Ben Fox Rubin 1/5/2015 4:56:17 AM
    Showing video of driving in UK in dark in bad weather.
  • Ben Fox Rubin 1/5/2015 4:56:41 AM
    Can see the speed camera -- Huang jokes that that's probably the most valuable part of the system.
  • Ben Fox Rubin 1/5/2015 4:56:53 AM
    He got a good laugh for that one.
  • Ben Fox Rubin 1/5/2015 4:57:22 AM
    Huang noted that it can classify all kinds of objects at once.
  • Ben Fox Rubin 1/5/2015 4:57:53 AM
    In congestion with braking traffic, it starts to get more triggers.
  • Ben Fox Rubin 1/5/2015 4:58:00 AM
    So can get to smoother stops.
  • Ben Fox Rubin 1/5/2015 4:58:20 AM
    Now the video is in Vegas.
  • Ben Fox Rubin 1/5/2015 4:58:34 AM
    Doing vehicle classification -- SUV, car, etc.
  • Ben Fox Rubin 1/5/2015 4:58:54 AM
    We're driving outside the Wynn, by the way.
  • James Martin 1/5/2015 4:58:55 AM
  • Ben Fox Rubin 1/5/2015 5:00:27 AM
    Huang said instead of having feature detectors for every type of vehicle, could train the system to learn these different vehicles. They "taught" it, he said.
  • Ben Fox Rubin 1/5/2015 5:01:03 AM
    This is how to use these powerful neural networks, which build on past learning.
  • Ben Fox Rubin 1/5/2015 5:01:37 AM
    Huang said the car learned all this stuff in just a few weeks.
  • Ben Fox Rubin 1/5/2015 5:02:22 AM
    Now switched to rear-facing cameras -- such as when the cops are following.
  • James Martin 1/5/2015 5:02:39 AM
  • Ben Fox Rubin 1/5/2015 5:02:51 AM
    Showing video of police with lights on trailing the car.
  • Ben Fox Rubin 1/5/2015 5:03:24 AM
    By the way, Nvidia engineer Mike Houston was presenting a lot of these videos.
  • James Martin 1/5/2015 5:04:24 AM
  • Ben Fox Rubin 1/5/2015 5:04:35 AM
    Now it's just Huang back on stage on his own. He sums up how Nvidia's deep learning technology can classify all kinds of objects.
  • Ben Fox Rubin 1/5/2015 5:04:55 AM
    "What comes out of the super-computer is a trained, deep neuro-net," Huang said.
  • Ben Fox Rubin 1/5/2015 5:05:14 AM
    The more objects this system processes the smarter it becomes.
  • Shara Tibken 1/5/2015 5:05:41 AM
    All of the demos we saw consumed less than 10% of Tegra X1's overall horsepower.
  • Shara Tibken 1/5/2015 5:05:46 AM
    (according to Huang)
Powered by ScribbleLive Content Marketing Software Platform
This week on CNET News See all