In computer graphics rendering means creating a digital image with the use of a computer - to be more exact the generation of a view of a virtual scene under an arbitrary viewing direction. Conventional methods use a scene description based on geometrical primitives to create a view. McMillan uses the term geometry-based rendering for these methods [MB95]. With those methods regular shaped 'artificial' objects can easily be described. The modeling of curved surfaces, however, is more complex. These surfaces have to be approximated with a polygonal mesh. Natural objects as trees or items of daily life as clothing must be laboriously constructed with thousands of single faces.
Recently there have been great advances in three-dimensional scanning methods, which promise to simplify the process of model building. However, this automated model creation also verifies our suspicion: the geometry of the real world is extremely complex [MB95]. Therefore the modeling process is a laborious and time consuming task. But not only the modeling, also displaying the scene is algorithmicaly complex and needs high computational power. To rotate or to move an object at interactive frame rates, it is necessary to generate at least 5 images/sec. For complex scenes this can only be done with dedicated hardware.
In contrast, image-based rendering uses pictures for describing the
scene. These images can be photographs, video takes or synthetic images. The
modeling process is replaced by image recording. New views are produced through
modification (interpolation, reprojection etc.) of existing input data. A great
advantage of image-based rendering is that the effort to display an
object is independent of the complexity of its internal representation. A tree
is represented by just as much pixels as a cube (assuming both are the same
virtual size). Thus, in an image-based rendering system the computational costs
of displaying new images are constant and independent of scene complexity. In
Table 1 the differences of the two methods are