Plane-sweep algorithm: various tools for computer vision (パターン認識・メディア理解)

Recent research on computer vision has made significant progress in 3d reconstruction and free-viewpoint video however most of these methods are not suited for real-time rendering. This paper presents a Video-Based Rendering method that provides online new viewpoints of the scene from a set of webcams. Our method follows a plane-sweep approach perfectly suited for GPU implementation. This paper mainly focuses on different implementations of this method for different purposes. Indeed, our basic plane-sweep technique uses 4 input cameras to create online new views. However this method can be modified to be used with up to 10 cameras or more by a camera selection process. This method can also be adapted to handle moving input cameras using a real-time camera calibration technique. Moreover, this method can easily render a depth map instead of a new view. Finally, we explain how the plane-sweep algorithm can be modified to create multiple new views simultaneously. This latter application is mainly designed for autostereoscopic displays application. Implementation and performance are detailed for all these plane-sweep methods.