Computer Vision News - November 2021

2 Summary Poster Presentation 30 Best of ICCV 2021 The work builds on recent advances in computer vision. It makes use of the deep neural networks that were designed for image classification a few years ago and capitalizes on that knowledge to adapt them to work for the problem of human mesh estimation . Neural networks do not work out of the box if you try to apply them to a new computer vision problem, especially a 3D computer vision problem. There is a need to incorporate domain-specific knowledge about 3D human mesh estimation. “ We had been inspired by the previous computer vision works in this area and that’s how we got the courage to work on this problem, ” Georgios reveals. “ We took inspiration from those and tried to combine and capitalize on previous ideas both from a mesh recovery direction and from the normalizing flows literature to get the best of both worlds. ” The work is one step towards achieving more accurate reconstruction from images, not only for 3D human mesh recovery, but much more generally, and people could apply this framework to other subfields in computer vision and machine learning. “We were very familiar with the human pose estimation task, but the challenge was to incorporate this probabilistic model in there. It’s different to read about them and what they do in principle. Making them work in practice takes a lot of experimentation.