A Progressive Rendering Algorithm Using an Adaptive Perceptually Based Image Metric Jean Philippe Farrugia and Bernard Péroche LIRIS, Université Claude Bernard Lyon 1, 8, boulevard Niels Bohr, 69622, Villeurbanne, France email: jpfarrug@bat710.univ-lyon1.fr bernard.peroche@liris.cnrs.fr April 22, 2004 Abstract In this paper, we propose to solve the global illumination problem through a progressive rendering method relying on an adaptive sampling of the image space. The refinement of this sample scheme is driven by an image metric based on a pow- erful vision model. A Delaunay triangulation of the sampled points is followed by a classification of these triangles into three classes. By interpolating each trian- gle according to the class it belongs to, we can obtain a high quality image by computing only a fraction of all the pixels and thus saving computation time. 1 Introduction Photorealism is a major goal to reach in computer graphics: the accurate simulation of complex environments needs to compute complete light transport between all parts of a virtual scene. This is possible by solving the rendering equation described by Kajiya in [Kaj86], and many solutions have already been proposed. Among all known meth- ods, we may quote hierarchical radiosity [HSA91], simple or bidirectional path tracing [LW93], radiance interpolation [War94b], photon maps [Jen96], light vectors [ZSP98], Metropolis light transport [VG97], density estimation [WHSG97], . . . All these meth- ods give realistic pictures, but despite the increase in computer performance, computa- tion times remain really long. Thus, a challenging problem is to dramatically reduce the computation time in order to be able to solve the global illumination problem at interac- tive rate. One way to evolve in this direction is to use a progressive rendering method: by carefully choosing a sample set, an approximate picture is generated. If the quality of this picture is not evaluated as sufficient, new samples may be selected to produce a new image and the process can be repeated until reaching the desired quality. The main problem with commonly used progressive methods is to decide where to cast these new samples. This usually relies on empirical criteria based more or less on statistics. We propose to replace these criteria with perceptual ones as they are more general and nat- ural than statistics based ones. Usually, statistical criteria take into account only one 1