Monday, December 01, 2014, 12:00pm - 07:00pm
Graduate Student, Rutgers University, Department of Computer Science
Copy to My Calendar (iCal)
Line drawings depict 3D shape using "minimal" information. Computer graphics algorithms have used different geometric surface features to define lines. While no line definition is universally superior to others (Cole09), the conditions under which some line definitions outperform others, are not well understood.
We examine the efficacy of two line definitions in conveying 3D shape. “Suggestive contours” are occlusions in nearby viewpoints (DeCarlo03). “Apparent ridges” are extrema of view-dependent curvature (Judd07). In our first experiment we look at how good of a predictor feature salience is. We expect that apparent ridges are more effective along sharper curvature extrema, while suggestive contours are more effective near more well-defined inflections. Thus, as a surface moves from a sharp triangle wave to a smooth sinusoidal wave, the better depiction should switch from apparent ridges to suggestive contours.
We find that for apparent ridges, feature salience is a good predictor of perceptual accuracy. Apparent ridges were also seen as sharper than the depicted surface for smoother objects. While the performance of suggestive contours was poor with objects of low salience, the performance with higher salience varied depending on their interaction with other geometric features, such as occluding contours, parabolic lines, and ridges/valleys.
Using these results, we created a classification scheme for suggestive contours based on their interactions with other geometric features. Suggestive contours a placed into one of five categories based on three yes/no questions. In a second experiment, used a gauge figure study to compare the accuracy of the percepts of suggestive contours from each category. We find that categories that answer yes to more of these questions perform better. This can help line-drawing generation algorithms select effective lines to depict 3D shape.