Towards Benchmark Data Generation for Feature Tracking in Scalar Fields

Emma Nilsson, Jonas Lukasczyk, Talha Bin Masood, Christoph Garth, Ingrid Hotz

View presentation:2022-10-17T21:41:00ZGMT-0600Change your timezone on the schedule page
2022-10-17T21:41:00Z
Exemplar figure, described by caption below
An example of a time-dependent 2D benchmark dataset for evaluating feature tracking in scalar fields. Four points are propagated through a vector field with a sink in the center during fifteen timesteps. In the upper row, the spatial embeddings of the points are shown with the scalar field and iso-contours. The scalar values in the field are calculated by taking the max of the Gaussian functions each point carries. Below the embeddings is the tracking graph of the four features, which merge in multiple steps into one feature.

The live footage of the talk, including the Q&A, can be viewed on the session page, TopoInVis: Session 2, Early Career Lightning Talks + Best Paper Awards .

Keywords

Human-centered computing—Visualization—Visualization design and evaluation methods; Human-centered computing—Visualization—Visualization application domains—Scientific visualization

Abstract

We describe a benchmark data generator for tracking methods for two- and three-dimensional time-dependent scalar fields. More and more topology-based tracking methods are presented in the visualization community, but the validation and evaluation of the tracking results are currently limited to qualitative visual approaches. We present a pipeline for creating different ground truth features that support evaluating tracking methods based on quantitative measures. In short, our approach randomly simulates a temporal point cloud with birth, death, split, merge, and continuation events, where the points are then used to derive a scalar field whose topological features correspond to the points. These scalar fields can be used as the input for different tracking methods, where the computed tracks can be compared against the ground truth feature evolution. This approach facilitates directly comparing the results of different tracking methods, independent of the initial feature characterization.