Optical Flow 2D and 3D for SciVis

Unsupervised learning of optical flow for scientific datasets Tasks: flow estimation, interpolation, time step selection

Datasets used:

Several ensembles from ETH Zurich (https://cgl.ethz.ch/research/visualization/data.php):

Fluid Simulation Ensemble for Machine Learning

Cylinder Flow with von Karman Vortex Street

Cylinder Flow Around Corners

Research Vessel Tangaroa

Droplet datasets (Uni Stuttgart):

Droplet 2D (link): The second dataset Drop Dynamics stems from a physical experiment to study the impact of a droplet with a film~\cite{geppert2016classification}. The captured experiment images are monochrome and have a resolution of 160x224. In total, there are 135K images from 1K members.

Droplet 3D (https://www.dropbox.com/s/yyz4pwf8g1mw6hc/drop.zip?dl=0): Generated by the Institute of Aerospace Thermodynamics in Stuttgart, a uniform resolution of 256x256x256 in Byte resolution (0 means air, 255 denotes fluid, there is nothing in-between; it stems from a two-phase flow simulation). Each time step is stored in one raw file in the order of (x,y,z), i.e, the first 256 elements depict the first row (the x-dimension), and the first 256x256 elements describe the first x-y slice.

Jets 3D dataset (https://www.dropbox.com/s/lduqpmd62rw7w6f/5Jets.7z?dl=0): It is a dataset with density and velocity fields, consisting of 2000 time steps (about 50G compressed, 80G uncompressed).

Synthetic datasets: Rectangle with high frequency texture, in 2D and 3D.