CW 557

Peter Vangorp, Timothy S. Condon, James A. Ferwerda, Kavita Bala, Roeland Schoukens, Philip Dutré
Visual equivalence in dynamic scenes

Abstract

Characterizing the fidelity of an image is a fundamental problem in computer graphics. Visible difference predictors are used to detect when images are perceptually indistinguishable. Visual equivalence introduces a new appearance-based standard for image fidelity, where two visibly different images of an object can still be visually equivalent as representations of the object's appearance if they convey the same shape, material and lighting properties. The original visual equivalence study focused on static scenes. In this paper, our goal is to understand and evaluate visual equivalence in the context of dynamic scenes.

We conducted a series of psychophysical experiments that explore visual equivalence for objects undergoing rotational and linear motion under a variety of illumination transformations. We demonstrate that equivalence continues to exist with motion for transformations that maintain some important properties. Specifically, they should not affect the sense of structured motion in the patterns of surface reflections. Based on these insights, we propose a new interpolation transformation that preserves some of these dynamic visual features.

report.pdf ( 15M) / mailto: P. Vangorp