Computer Vision News - December 2020

246 a mess of points. You have to be able to create coherence across those individual messes of 3D points. That’s something that we learned to do automatically. That made it possible to grab 3 scans from any source and take statistics from all of these different 3D scans and create this coherence. We were the first ones to do that in a way that was easy to use. We also learned the muscle deformations. That was the one thing that nobody else had at that point. Still, nobody was able to do it as well. I’m not taking credit for it! [ laughs ] So many peoples’ work has gone into developing this system: ten, twenty different people who have come and gone. It has been years of work. That’s why it’s very special. Any other company that has enough resources can collect the data. Curating that data and going through and creating this coherence (which is also creating clean 3D datasets and then creating statistical models out of it), that has taken years and a lot of expertise. Still, there are many people out there who are not using this 3D model system. So we have developed these 3D models, and now there are people who can build on top of that who can create systems that can identify bodies in images. In the background, they still have a 3Dmodel, and that’s what people use. They use our model. This is used in methods for automatically cleaning 3D scans, automatically detecting bodies in images, automatically detecting RGBD data. The application grew from there. Now, people are also using this same model to create lots of synthetic data, so that they can train AI systems for autonomous driving. They can create synthetic routes for autonomous driving systems to learn from. People use this model because it can create the most realistic looking people, realistic looking body shapes and deformations from motion. That’s why people come to us! We forgot to mention one really important fact about this venture. You are not doing it alone. You are working with Michael Black! Can you tell us about Michael’s influence on this project? This whole idea of developing a 3D body model was something that Michael and his group were already working on, when I joined Max Planck. They were using an older version of the model that was called the SCAPE model. They created their own version of it using their own data. When I joined, we started developing applications around it using this older version. That’s when Women in Computer Vision