Abstract
Image-based rendering generates images from one or more known images without explicit geometric representation of objects. This paper presents a derivation of the pixel transformations for three-dimensional image warping, a popular form of image-based rendering. The derivation is based entirely on traditional three-dimensional graphics terminology, using only 4 × 4 matrices. The transformations are optimized based on the fact that transformations of adjacent pixels share much common computation. Derivatives and frame-to-frame coherence are also discussed.