Comparison metrics for the output of a mixed fixed point and floating point graphics pipeline

We propose to provide an analysis of some already existing and newly proposed metrics used for comparing altered images. The goal is to be able to accurately and objectively evaluate the alteration produced by an inaccurate, hybrid graphics pipeline that uses multiple real number formats relative to a fully accurate one. In order to obtain this, we have modified our own OpenGL ES 1.1 software implementation so that it could operate with different representations of real numbers. We fed this hybrid pipeline with about 2500 3D objects and rendered them in 180 frames for each object, in different positions, using a combination of real number formats across the pipeline. By using this setup we have obtained about 450000 reference images and about 2700000 altered ones. Finally, we obtained the differences between the two sets of images and tried to analyze them by using different metrics so that we could distinguish between acceptable altered images and non acceptable ones.