ECCV 2016 Daily - Tuesday

cloud. Then we fit some kind of intermediate model in those which helps us to create new points and track new tracks. ECCV Daily: Why did you choose this kind of project? What fascinates you in this predictive model? Karel: It first started as a tracking project when we were trying to follow some objects and find the trajectory in 2D. I have some interest in 3D geometry and 3D computer vision so I put these two together. I started working on the tracking and reconstructions simultaneously. Then, this non-rigid reconstruction naturally followed. ECCV Daily: What is the practical application of your model in the real world? Karel: You could use it if you wanted to manipulate some object to get its shape. One example is the use of any kind of video recording from a TV or from the internet. You could reconstruct a public performance of someone speaking. You could reconstruct the model of their head to be non-rigid. You could choose an expression. Of course, it wouldn’t be as good as if you had a face specific model, but in this case the advantage is that it’s very generic. It could work on any kind of object without prior knowledge. ECCV Daily: What is the next step? Karel: I finished working on this when I finished my PhD. Now I am working on a different project: collecting driving statistics from dashboard cameras in cars. I would like to use this opportunity to recommend the use of blendshapes, which are used quite often in computer graphics but not very much in computer vision, and I think they should be used more. In non-rigid structure from motion, you have this instantaneous time varying shape which is expressed as a linear combination of some basic shapes. In computer graphics, you often have these blendshapes when each of these basic shapes would be Presentations 9 ECCV Daily: Tuesday one extreme of the non-rigid deformation. Then you would interpolate between those. They are all meaningful and extreme, but still possible variations of that shape. In computer vision, often when people do non-rigid structure from motion, this is not used. There’s nothing enforcing these basic shapes to be these extreme, but still possible variations, and I think it’s useful to have them as such.