High Dynamic Range Pixels

Christophe Schlick Laboratoire Bordelais de Recherche en Informatique Talence, FRANCE We present a picture encoding technique for RGB color images that uses nonlinear quantization of pixel values to achieve a much higher dynamic range than the usual 256 values per primary color. The encoding uses only 24 bits per pixel. } Introduction } Global illumination algorithms (radiosity, Monte Carlo ray tracing) usually generate pictures with much higher dynamic ranges than direct illumination ones: a factor of 2000 between the highest and the lowest intensity values is common and a factor of 30000 is sometimes reached with special illumination cases (Chiu et al. 1993). Therefore traditional 24-bit picture representations with a dynamic range of 256 are inadequate. Moreover, when applying gamma correction, even pure 24-bit pictures can create color bandings for low values (Hall 1989) showing that a greater dynamic range is needed. Ward has proposed a technique (Ward 1991) in which pixels values are coded with a oating point format. Four bytes per pixel are used (one byte for a common exponent and one byte for the mantissa of each primary color) which provides a huge dynamic range of 2256, more than enough for rendering purposes. Nevertheless, that method has several weak points. First, the memory cost is increased by a third. Second, the coding/decoding routines make heavily use of the math routines frexp and ldexp which are typically slow (though more e cient implementations can be found). Third, using a common exponent for R, G and B favours the dominant color and can lead to color shifts (only the dominant component uses a full 8-bit mantissa, the two others have can have several leading zeros in their mantissa in order to compensate the inadequate common exponent). Moreover, oating point pixels only solve half the problem: if they enable storage and manipulation of high dynamic range pictures, mapping these pictures to the 256 values allowed by the typical frame bu er is still unclear. Logarithmic quantization has been proposed as a remedy to these problems (an overview as well as some implementation issues can be found in (Blinn 1989)). The technique is based on the observation that when quantizing pixel intensities, one should pay more attention when encoding low values than high ones. There are at least two