CVPR Daily - Thursday

There’s much work on how to generalize not only to new debris but to imaging conditions. We don’t have images from space of the debris that we’re trying to capture, so we need to be able to train our deep networks or whatever machine learning model we have on synthetic images, but they need to work in space. The other aspect is that we don’t have access to large clusters of GPUs in space. We have FPGAs and small CPUs. We need networks and machine learning models that work in real-time on minimal computer vision. There’s the whole aspect of network compression, pruning, and quantization to make things work in real-time. When did you graduate with your PhD? I graduated in early 2009. What has happened since then? Many things! First, I went to do a postdoc at UC Berkeley with Trevor Darrell. I was there for a year and a half, working on 3D shape and non- rigid reconstruction. Then I moved to TTI in Chicago. I continued working on a similar topic. A bit of 3D human pose estimation as well. Then I moved to Australia. I had a senior researcher position there. That’s where I started broadening my scope, working on Riemannian geometry for computer vision, visual recognition, segmentation, and many different problems. Is that where you took the famous photograph of yourself with the cowboy hat? Exactly! [ laughs ] It’s made of kangaroo leather! I was in Australia for three and a half years. My first daughter was born in Australia. My wife is Swiss, so we decided to move back to Switzerland to have the family around. That’s when I had the opportunity to join CVLab with a permanent position at EPFL. Tell us something about EPFL that we don’t know. The faculty that I’m in, the computer science and communication systems faculty, has an interesting program for PhD students. We have a way of recruiting PhD students such that they can get a fellowship for the first 14 DAILY CVPR Thursday Exclusive Interview with…