Computer Vision News - July 2022

BEST OF CVPR 14 CVPR Poster Presentation interpretable models .” Finally, we tell Moritz that Bernt Schiele , his supervisor, has told us to give a look at this work. “This makes me so happy that he’s supportive of this paper and took the chance to recommend it for CVPR Daily!” he smiles. “Working with Bernt is special because of the close supervision, discussions, and feedback. It helps me refine my argumentation. He’s also very fast! Both in terms of speaking, as many people will know, and in terms of understanding and connecting the dots.” Is it true he can make the best out of the people he works with? “Yes, that fits him very well. He knows where the value lies.” “I didn’t manage to go to many invited talks because there was such a lot of interest in my poster!” he laughs. “I got very positive feedback, which encourages me.” Why does he think his work has proved so popular, particularly picked out from the competition for a coveted CVPR poster slot? “I think it’s because typically, there has been this trade-off between accuracy and interpretability,” he answers. “We’ve been getting better and better in classification performance, but we’ve been stuck when it comes to understanding exactly how the models perform their tasks. We show in our paper that if we design the models the right way, we can maintain accuracy but build inherently