Model-based video coding with generic 2D and 3D models

D extraction from video gives a representation adapted to low bitrate coding and provides enhanced functionalities such as 3D cues for augmented reality and interactive na- vigation in photo-realistic environments. But for degenera- ted motions of camera, like pure rotation, 3D information can not be retrieved. In this article we propose an original representation based on a hybrid 2D/3D models stream. This approach provides a modelling for all sequences in- cluding those with rotations. The sequence is divided into portions and for each one the motion of the camera is iden- tified. Depending on the type of motion a 3D model, a cy- lindrical mosaic or a spherical mosaic is extracted. They are constructed in order to be suitable for an homogeneous visualization process. Results are shown for synthetic and real video sequences.