Computer Vision News - July 2021

Fabio Tosi / Yiyi Liao 33 Best of CVPR 2021 Fabio, who was already an expert on stereo matching when he joined the team, adds : “ I never imagined a general framework capable of predicting depth at arbitrary spatial resolution while keeping precise and sharp estimates at object boundaries. I’m still impressed by the elegant formulation and the results achieved. ” In terms of next steps, with the output representation being so general, they plan to extend it to optical flow and self-supervised depth estimation. They are also interested in 3D reconstruction, which requires depth fusion of multiple images. Thinking about challenges, the team faced one big one in particular, which was trying to work together during a pandemic! With Fabio in Bologna and other team members including Yiyi and Carolin Schmitt in Tübingen, they have conducted the entire project remotely and are yet to actually meet in person. “ We organized it quite well even with this remote collaboration! ” Yiyi smiles. “ We had great teamwork and worked together well to overcome any challenges. For example, at the beginning we had this output representation and we were thinking how we should train it. We came up with all different loss functions, like how to disentangle the foreground and background, which requires some extra supervision. In the end, what’s surprising is that actually with a simple likelihood- based loss function then it works perfectly without extra supervision as we were proposing. ” The initial idea for this work came from Professor Andreas Geiger. This is the second time we have spoken to somebody from his team, following our interview wit h Despoina Paschalidou at CVPR last year. “ He is an excellent supervisor! ” Yiyi beams. “ He gives very insightful feedback, both at the high-level and at a technical level. Whatever problem you are having, he is able to point you to the correct solution. I think he’s really supportive – and I know everyone in our group agrees! ” Fabio adds : “ Working with Andreas and his team was really exciting. The way they work and deal with problems is impressive. They opened my mind and taught me how to tackle research in a very different and effective way. I am extremely grateful to Andreas for this and for giving me the opportunity to join his incredible group. ” To find out more about Fabio and Yiyi’s work, you can view their presentation and PDF poster anytime, and take part in their live Q&A during Paper Session Seven today (Wednesday) at 11:00-13:30 EDT.