Efficient 3-D Scene Visualization by Image Extrapolation

Tomás Werner, Tomás Pajdla, Václav Hlavác

Abstract: Image-based scene representation is believed to be an alternative to the 3-D model reconstruction and rendering. In attempt to compare generality of image-based and model-based approaches we argue that it is plausible to distinguish three approaches to 3-D scene visualization: image interpolation, image extrapolation, and 3-D model reconstruction and rendering. We advocate that image extrapolation is a useful trade-off between simple but limited interpolation and general but difficult 3-D model reconstruction and rendering. Image extrapolation is able to visualize correctly the part of a 3-D scene that is visible from two reference images. In fact, it is equivalent to reconstructing a projective 3-D model from two reference images and rendering it. In the second part of the work, we present an algorithm for rendering a projective model. Our approach is more efficient than the ray-tracing-like algorithm by Laveau and Faugeras [6]. We show that visibility can be solved by z-buffering, and that virtual images can be synthesized by transferring triangles from a reference image via a homography or an affinity. Such algorithms are often supported by hardware on graphics work stations, which makes a step towards the real-time synthesis. The results are presented for real scenes.