[MUSIC PLAYING] KILLIAN WEINBERGER: One of the essential problems in self-driving cars is to identify objects around the car-- for example, other cars, pedestrians, et cetera. The goal is to basically put bounding boxes around objects so that we can circumvent them and do that as precisely as possible. LIDAR is very precise, but a drawback of LIDAR is that it's very expensive. And it sticks out in the top of the car, so it induces wind resistance. The question is can we also do 3D object detection without a LIDAR?
And the way you would do this is by using stereo camera. So the same way humans have two eyes, and you can see depth with our two eyes, you could also have two cameras in the car. And each object you see, how far off is in the right or left camera that tells you how far away it is.
And cameras are dirt cheap. They're just a few bucks. And you install them on both sides of the car, and you put it behind the windshield, so they're not even inducing any wind resistance.
Because now there's so many people doing so many things in computer science or close to computer science, this is a great example of collaboration. Bharath is a computer vision expert and Mark is a robotics expert at SBZ-- all these different experts within three minutes' walking distance around my office.
I hope this project contributes significantly to making self-driving cars cheaper and more feasible and more reliable-- cheaper, because you don't need LIDAR anymore, more reliable because now you have redundancy. You have two different sensors. People thought the difference was the difference in accuracy. But the difference in accuracy is actually not very big. It's actually the perspective that makes a difference.
The trick is to take the data that you have, convert it to 3D information, then change the perspective. And then, analyze it from above. You can identify the cars with a very, very high accuracy, very close to what we get with LIDAR, just at a fraction of the cost.
We've received your request
You will be notified by email when the transcript and captions are available. The process may take up to 5 business days. Please contact email@example.com if you have any questions about this request.
Cornell researchers have discovered a relatively simple, low-cost method that allows autonomous cars to detect 3D objects with high accuracy. Kilian Weinberger, associate professor of computer science and senior author of the paper, "Pseudo-LiDAR from Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving," discusses the group's research.