Extracting flow features via supervised streamline segmentation. (November 2015)

• Record Type: Journal Article
• Authors: Li, Yifei; Wang, Chaoli; Shene, Ching-Kuang
• Is Part Of: Computers & graphics. Volume 52(2015)
• Page Start: 79
• Page End: 92
• Publication Date: 2015-11
• Subjects: Flow visualization -- Flow feature extraction -- Streamline segmentation -- Support vector machine; Computer graphics -- Periodicals; 006.6
• DOI: 10.1016/j.cag.2015.06.003
• Languages: English
• ISSNs: 0097-8493

Abstract

Abstract: Effective flow feature extraction enables users to explore complex flow fields by reducing visual clutter. Existing methods usually use streamline segmentation as a preprocessing step for feature extraction. In our work, features are directly extracted as a result of streamline segmentation. In order to achieve this, we first ask users to specify desired features by manually segmenting a few streamlines from a flow field. Users only need to pick the segmentation points (i.e., positive examples) along a streamline, remaining points will be used as negative examples. Next we compute multiscale features for each positive/negative example and feed them into a binary support vector machine (SVM) trainer. The trained classifier is then used to segment all the streamlines in a flow field. Finally, the segments are clustered based on their shape similarities. Our experiment shows that very good segmentation results can be obtained with only a small number of streamlines to be segmented by users for each data set. We also propose a novel heuristic based on the minimum bounding ellipsoid volume to help determine where to segment a streamline. Abstract : Graphical abstract: Abstract : Highlights: We proposed a supervised segmentation algorithm to extract flow features. We propose an effective heuristic which captures how human beings segment streamlines. We illustrate the utility of our streamline segmentation method using both synthesized and real-simulation data sets. We compare our method with the state-of-the-art to show our method performs much better.