Abstract

Optical flow algorithms were originally developed for estimating the motion of opaque objects. In the case of semi-transparent fluids, the object is a volumetric distribution that exhibits motion in three dimensions. In principle, adapting optical flow algorithms to 3D flow seems simple: we still have a 2D image to work with, representing a projection through a 3D volumetric object. And now we need only consider the velocity estimate as being an average through the object along the line of sight, instead of the velocity at a single point. Unfortunately, however, this simple step creates flaws in the algorithm that are easily overlooked.

© 2017 Japan Society of Applied Physics, Optical Society of America