Suppose you have a sequence of images and you want to infer the “image flow” of the motions depicted in the sequence. These motions might be the result of activity by various objects in the scene or they might be motions induced by the movement of the observer, so-called egomotion.

Current technology allows the creation of larger images, larger in the sense of higher resolution, which necessitates some form of sampling of the image spaces to reduce the amount of computation required. One way to do this sampling is to borrow from biological vision systems and use a log-polar representation, in which the density of the samples decreases as one moves away from the focal point. However, as the authors of this paper point out, for a number of potential applications, such as an autonomous vehicle with a forward-facing vision sensor, this means that the immediate foreground is less densely sampled than the region around the vanishing point at which the sensor is directed.

The authors propose a sampling technique, reverse log-polar, in which sampling is densest at the edge of the (circular) vision field. The authors justify using a polar coordinate system for the motion flow computation problem on the basis of image motion statistics. They compare the accuracy of log-polar and reverse log-polar sampling using both synthetic and real-world sequences. The average angular error is used to determine the accuracy compared to the ground truth optical flow.

The experimental results show that reverse log-polar representation is superior to log-polar, at least for the forward-facing sensor systems considered. It would be interesting to consider systems that use a mixture of sampling techniques.