Predicting and Evaluating Saliency for Simplified Polygonal Models

Abstract

In this paper, we consider the problem of determining feature saliency for three-dimensional (3D) objects and describe a series of experiments that examined if salient features exist and can be predicted in advance. We attempt to determine salient features by using an eye-tracking device to capture human gaze data and then investigate if the visual fidelity of simplified polygonal models can be improved by emphasizing the detail of salient features identified in this way. To try to evaluate the visual fidelity of the simplified models, a set of naming time, matching time, and forced-choice preference experiments were carried out. We found that perceptually weighted simplification led to a significant increase in visual fidelity for the lower levels of detail (LOD) of the natural objects, but that for the man-made artifacts the opposite was true. We, therefore, conclude that visually prominent features may be predicted in this way for natural objects, but our results show that saliency prediction for synthetic objects is more difficult, perhaps because it is more strongly affected by task. As a further step we carried out some confirmation experiments to examine if the prominent features found during the saliency experiment were actually the features focused upon during the naming, matching, and forced-choice preference tasks. Results demonstrated that the heads of natural objects received a significant amount of attention, especially during the naming task. We hope that our results will lead to new insights into the nature of saliency in 3D graphics.