Computer Vision News - July 2020

them as well and get the shape of these nasty things! We’ve also been working on combining deep SDFs with explicit surface representations. With all my baggage, I like 3D meshes. They have their advantages. There are cases in which you actually need them. On the other hand, it’s been shown that SDFs are great so that’s something we’ve been working on. We’re working right now to combine the two and attain the holy grail of end-to-end differentiability. We are praying to those gods every morning. How many people are you in the lab? W hen we have a full Zoom meeting, we are 33! I’ve been looking through your alumni and a few star names jumped out at me. One in particular that caught my eye was awesome Raquel Urtasun. Do you recall what she was like as a PhD student? I knew from day one that she would go very far. We still talk at regular intervals – not as much as I would like, as she is extremely busy. In my lectures this year, I used some slides from her thesis about Gaussian processes. Another thing Raquel was working on is human pose estimation. We’ve done a lot of this and we still are. Like the 3D surfaces, it’s gone deep. However, the important fact is that, for a given human behaviour, all poses exist on some manifold in this very high-dimensional space. It was true then and it’s still true. In Raquel’s time, we modelled this with Gaussian processes, and now we tend to use autoencoders. The machinery has changed, but the underlying idea not so much. You may have noticed Mathieu Salzmann is also in my lab. He is a good friend of Raquel. So, from time to time, we still talk about Gaussian processes with a hint of nostalgia. Pascal Fua 27 Past PhD student, now Associate Professor Raquel Urtasun "It’s my students I have to thank for not drowning!" Best of CVPR 2020