ECCV 2016 Daily - Thursday

ECCV Daily: What is the work you are presenting? Matthias: We want to use a commodity depth sensor or RGB-D so that we have color images and depth images. We want to jointly reconstruct the geometry and motion of the scene. Previous approaches mostly rely on templates. They have a pre- scanned template mesh or other representations of the scene like a human body or face. Do you know the famous face project from Matthias Niessner where they have a prescanned representation of the face? We wanted to do something that reconstructs from scratch. Compared to other previous approaches, we wanted to improve the quality. Since we were probably not able to reconstruct them, So we incorporated color features as well. ECCV Daily: Is this the main innovation in your work? Matthias: Well, it’s two-fold, I would say. One side is that we incorporated color features in the real time reconstruction. Also, we had a higher resolution deformation field that allowed us to represent motion and deformation at a higher accuracy. ECCV Daily: What was challenging about the project? Matthias: It was very challenging to get everything to work together especially under the real-time constraint. You have to solve it in 30 milliseconds. You have to do the whole reconstruction part, correspondence estimations, SIFT matching, optimization of a non- linear problem, high-dimensional non-linear problem. You have to work very hard to get it running on the GPU. VolumeDeform: Real-time Volumetric Non-rigid Reconstruction 14 ECCV Daily : Thursday Presentations Matthias Innmann is a PhD candidate working in the computer graphics group at the university of Erlangen-Nuremberg “It was very challenging to get everything to work together especially under the real-time constraint” Keep reading us also after ECCV 2016. Subscribe now for free at Computer Vision News (click here)