This paper presents a semi-automatic streamline segmentation algorithm, which can be used to identify user-specified features from a streamline soup. The pipeline first clusters the streamlines into groups, and the user needs to manually pick a (or some) valid segmentation point(s) for a representative streamline in each group for training purposes. Both positive and negative samples are generated automatically based on these user-selected points. In the end, a support vector machine (SVM) is trained and then used to generate the final segmentation results. There is not a rich explanation about why a specific algorithm is used, but it seems intuitive enough and little effort is required to fit other algorithms into the proposed framework.

Geometric processing is not expected to be a neat and well-defined task. The proposed algorithm needs a lot of tweaks here and there to work, and those are where we need to pay close attention to, for example, neighborhood selection, dealing with highly imbalanced training data, and so on. The authors explain these details very well. The meat of this paper is section 3.2, that is, the selected features. The bounding ellipsoid feature is especially interesting; the authors emphasize its significance in table 2. In general, it is an audacious effort to define what should be seen as a feature for a curve. The authors imply this at the end of section 2: “We have not found any rules from cognitive science [that] can help us segment a 3D curve (for example, streamlines).” Thus, if their methods prove to be well accepted by a large group of different readers, it may help fill the gap in cognitive science in a way.

The computation is very efficient on a personal computer for small datasets. For medium and large datasets, a parallel version is needed, as indicated in the future work section.