ECCV 2016 Daily - Wednesday

We spoke with Jean-Sebastien Franco , Associate Professor at Ensimag and INRIA Grenoble Rhône-Alpes , who works with a team interested in retrieving appearance representations for full- sequences of 4D recordings: people that are recorded by multiple camera setups. The project seeks to capture their performance, their geometry, and their appearance (or texture). Franco works with Adnane Boukhayma , a PhD Student, who is the main author, another PhD student, Vagia Tsiminaki , and Edmond Boyer , who leads the group. The novelty of their work is mainly the texture in that they are trying to capture the variability of the texture of the person being captured over the whole sequence. In existing works, textures have been constructed per frame or for small time frame window. Here, they are trying to get a presentation that factors all of the variability for multiple sequences for one sequence, but also for multiple sequences of the same person. It is difficult when you involve a lot of cameras. One of their earlier works computed a high resolution texture in all of these cameras in a single high resolution texture. Getting as much detail as possible for one time instant already proved challenging. Now that the team went through this step, one of the things that make it difficult when going over all of the sequence is the assumptions when you have one time frame. With one time stamp, the lighting condition and the geometry doesn’t change. However, when you go over the full sequence, the frames have micro variations in the geometry along with lighting change because of changes in the body position. The texture is not actually being observed over the same angle with the light sources. All of this variability can’t be captured in just one texture. It requires some kind of representation to capture the variability over all of the sequences to capture these kind of variations. Eigen Appearance Maps of Dynamic Shapes 8 ECCV Daily : Wednesday Presentations