CVPR Daily - Wednesday

Wednesday 5 Why do you have such a strong and active presence at CVPR? You are also a diamond sponsor of the event. Computer vision and machine learning are two very important parts of NVIDIA nowadays, for obvious reasons. Computer vision due to self-driving cars and robotics - the two dominant application areas of computer vision. Machine learning in general, of course, is very important. How many papers are you presenting at CVPR this year? We have been very successful this year, with 14 full papers at CVPR. 11 of them are in collaboration with my group. You don’t want to talk about all 14 of them? Yes. You have 13 seconds for each! [ both laugh ] There’s the work which we call Super SloMo, a method for taking standard footage that’s recorded, say at 30 frames per second, and we can slow it down to an arbitrary rate using a neural net. It works very well. We have trained it in a self-supervised manner on lots of videos. Now we can insert or hallucinate intermediate frames to slow down a video; that works really nicely. What problem in the real world will that solve? One of the use cases is aesthetics – people like to look at things in slow motion. There are also some professional use cases. For example, if you are a professional ballerina or athlete and you want to see every nuance of your form, you could do that with this method by slowing it down and looking at it in slow motion. Another paper, in collaboration with an intern from UMass Amherst, is called SPLATNet. That is also very interesting and has won the best honourable mention award. To generalise, it is about dealing with sparse, high- dimensional data. In particular, we looked at point clouds and how you process point clouds. Looking at processing not just the point clouds themselves, but also point clouds together with images. Say, you have a LiDAR scan as well as RGB images, you can work or process them jointly. For instance, to do semantic segmentation, which is really quite neat. Can you tell us something nice about NVIDIA that the public doesn’t know? [ laughs ] NVIDIA is a very open company internally, which is great: we share a lot of information across teams. It is very collaborative, which is probably not visible from the outside, but it is a very fun place to work as a result. That sounds awesome. What can you share with us about what’s coming up at NVIDIA in the future? You can expect to see research expand. We are eager to push more on the research front and be visible at all the major conferences. It is important to us to be part of the community and really give back to the community as well. I am very happy to be at CVPR and it is great to see so many people attending and there being such a strong interest in computer vision. The research that people are doing and showing here is great to see. Jan Kautz - NVIDIA