Inferring Reflectance Functions from Wavelet Noise

Abstract

This paper presents a novel method for acquiring a wavelet representation of the reflectance field of real objects. Key to our method is the use of wavelet noise illumination to infer a reflectance function for each pixel. Due to their stochastic nature, these wavelet noise patterns enable to trade off the number of recorded photographs for the quality of the computed reflectance functions. Additionally, each wavelet noise pattern affects all pixels in a recorded photograph, independently of the underlying material properties in the scene. Consequently, each recorded photograph contributes additional information to the reflectance field computation. The presented method consists of three steps. First, a fixed number of photographs are recorded of the scene lit by a series of wavelet noise patterns emitted from a CRT monitor. Next, for each pixel a reflectance function is computed offline, by identifying the important wavelet coefficients for the pixel's reflectance function. The coefficients are computed by solving a linear least squares problem. Finally, once all reflectance functions are computed, a novel image of the scene can be composited with arbitrary incident illumination. The method can be used for both image-based relighting and environment matting.