A comparison of scene flow estimation paradigms

Abstract

This paper presents a comparison between two core paradigms for computing scene flow from multi-view videos of dynamic scenes. In both approaches, shape and motion estimation are decoupled, in accordance to a large segment of the relevant literature. The first approach is faster and considers only one optical flow field and the depth difference between pixels in consecutive frames to generate a dense scene flow estimate. The second approach is more robust to outliers by considering multiple optical flow fields to generate scene flow. Our goal is to compare the isolated fundamental scene flow estimation methods, without using any post-processing, or optimization. We assess the accuracy of the two methods performing two tests: an optical flow prediction, and a future image prediction, both on a novel view. This is the first quantitative evaluation of scene flow estimation on real imagery of dynamic scenes, in absence of ground truth data.