Database construction & recognition for multi-view face

We present data collection and recognition experiment focused on multi-view face recognition/descriptor. Many face databases and face recognition systems have been constructed and experimented in terms of various illumination, time, poses, or expressions. However none of databases yet satisfies a large variation of poses to study systematic 3D human face information, which results unsatisfactory success rate for the posed face recognition while many quite satisfactory frontal view reconstructions have been shown. It is due to the difficulty of data collection of facial images to satisfy the large variation of poses to fully represent the 3D characteristic of human faces. We show two possible multi-view face data collection either using rendering of 3D models or using a video camera. We also illustrate our approach to build a face descriptor containing 3D information of human face using multiview concepts. This multi-view face recognition descriptor is a 3D face descriptor which takes systematic extension of 2D face descriptor using the concept how much powerful a view influences over nearby views, so called as "quasi-view" size.