Optical Flow for Rigid Multi-Motion Scenes

We observe that in many applications, the motion present in a scene is well characterized by a small number of (rigid) motion hypotheses. Based on this observation, we present rigid multi-motion optical flow (RMM). By restricting flow to one of several motion hypotheses, RMM produces more accurate optical flow than arbitrary motion models. We evaluate an algorithm based on RMM on a novel synthetic dataset, consisting of 12 photo-realistically rendered scenes containing rigid vehicular motion and a corresponding, exact, ground truth. On this dataset, we demonstrate a substantial advantage of RMM over general-purpose algorithms: going from 36% outliers with the DiscreteFlow algorithm, to 26% with ours, with a mean error reduction from 8.4px to 6.9px. We also perform qualitative evaluation on real-world imagery from traffic cameras.