Image-based rendering via the standard graphics pipeline

Ullman and Basri (1991) have shown that a 3D scene can be implicitly represented as a linear combination of two or more basis-images. We outline a simple visualisation system which uses this result to provide interactive viewpoint control. Moreover, we show that each novel view can be rendered without significant departure from the standard graphics pipeline. Essentially, only texture-map, z-buffer and accumulation-buffer facilities are required. The need for dense feature-correspondence is also avoided. It is desirable for image-based rendering to follow the standard pipeline, in order to exploit hardware-support for existing graphics APIs. As an example, we present the results of an OpenGL implementation of the system. Our observations are relevant to augmented reality, visualisation, animation and video coding applications. We also consider the possible use of image-based representations in network-graphics.