II.5 - REAL PIXELS

Abstract

This chapter analyzes the concept of real pixels. It discusses a floating point format that only requires 32 bits per pixel and is completely portable between machine architectures. The idea is using an 8-bit mantissa for each primary and following it with a single 8-bit exponent. In most floating point formats, the mantissa is normalized to lie between .5 and 1. Because this format uses the same exponent for three mantissas, only the largest value is guaranteed this normalization, and the other two may be less than .5. It appears that this format favors the largest primary value at the expense of accuracy in the other two primaries. This is true, but it is also true that the largest value dominates the displayed pixel color so that the other primaries become less noticeable. The 32-bit real pixel format presented in the chapter preserves the bits that are most significant, which is the general goal of any floating point format. The chapter highlights that besides the ability to perform more general image processing without losing accuracy, real pixels are great for radiosity and other lighting simulation programs, because the results can be evaluated numerically well outside the dynamic range of the display.