High dynamic range preserving compression of light fields and reflectance fields

Surface structures at meso- and micro-scale are almost impossible to convincingly reproduce with analytical BRDFs. Therefore, image-based methods like light fields, surface light fields, reflectance fields and bidirectional texture functions became widely accepted to represent spatially nonuniform surfaces. For all of these techniques a set of input photographs from varying view and/or light directions is taken that usually by far exceeds the available graphics memory. The recent development of HDR photography additionally increased the amount of data generated by current acquisition systems since every image needs to be stored as an array of floating point numbers. Furthermore, statistical compression methods -- like principal component analysis (PCA) -- that are commonly used for compression are optimal for linearly distributed values and thus cannot handle the high dynamic range radiance values appropriately. In this paper, we address both of these problems introduced by the acquisition of high dynamic range light and reflectance fields. Instead of directly compressing the radiance data with a truncated PCA, a non-linear transformation is applied to input values in advance to assure an almost uniform distribution. This does not only significantly improve the approximation quality after an arbitrary tone mapping operator is applied to the reconstructed HDR images, but also allows to efficiently quantize the principal components and even apply hardware-supported texture compression without much further loss of quality. Thus, in addition to the improved visual quality, the storage requirements are reduced by more than an order of magnitude.