Evaluate Neighbor Search for Curve-based Vector Field Processing

Abstract: Curve-based representations, particularly integral curves, are often used to represent large-scale computational fluid dynamic simulations. Processing and analyzing curve-based vector field data sets often involves searching for neighboring segments given a query point or curve segment. However, because the original flow behavior may not be fully represented by the set of integral curves and the input integral curves may not be evenly distributed in space, popular neighbor search strategies often return skewed and redundant neighboring segments. Yet, there is a lack of systematic and comprehensive research on how different configurations of neighboring segments returned by specific neighbor search strategies affect subsequent tasks. To fill this gap, this study evaluates the performance of two popular neighbor search strategies combined with different distance metrics on a point-based vector field reconstruction task and a segment saliency estimation using input integral curves. A large number of reconstruction tests and saliency calculations are conducted for the study. To characterize the configurations of neighboring segments for an effective comparison of different search strategies, a number of measures, like average neighbor distance and uniformity, are proposed. Our study leads to a few observations that partially confirm our expectations about the ideal configurations of a neighborhood while revealing additional findings that were overlooked by the community.