CVPR Daily - Wednesday

Pascal, could you tell our readers what your work is about? It’s about figuring out how the human mind works. That’s what computer vision is at its core. That is a big definition. Can you be a little more specific? Yes, sorry, that was a bit of a cop out! But it’s true, and actually, with deep learning, we are very far away from that because we have something that works nicely, but we don’t really understand why. We’ve deviated from that goal quite severely. You have published or co-published more than 300 papers to date. That is a staggering achievement. What are the big changes that you have seen over that time? We’ve been modelling 3D surfaces for many years now and are still at it. It’s interesting because 10-15 years ago we were doing surfacemeshes with classical correspondences and photogrammetry. Now, in 2020, we’re still looking at surface meshes, but they’ve become deep! A lot of the recent work is about, how do you bring all of this essentially old knowledge into the modern, deep world. In 3D modelling last year there was a small revolution with signed distance functions (SDFs), which have really had an impact, but this is not a new idea. I’m old enough to remember all the excitement around Sethian and Osher and implicit surfaces. We were all very impressed. It was very cool, and still is, but one of the difficulties was that you had to store these huge cubes of data which wasn’t very practical. Now we can replace the cube of data with a deep net, and it’s interesting because it’s a mix of the old and the new. Well, I would argue it’s a mix of the old and the old because the implicit surfaces are old, and the networks are old as well. What’s new is to mix them in a clever way. What have you been working on in your lab recently? We’ve recently worked on combining explicit and implicit surface representations to segment biomedical imagery, which means segmenting synapses to be precise. We do a lot of that in the lab. In fact, if I could get my hands on some coronavirus EM pictures, I would very gladly segment DAILY Wednesday Pascal Fua 8 Pascal Fua