View-Consistent 3D Scene Flow Estimation over Multiple Frames

We propose a method to recover dense 3D scene flow from stereo video. The method estimates the depth and 3D motion field of a dynamic scene from multiple consecutive frames in a sliding temporal window, such that the estimate is consistent across both viewpoints of all frames within the window. The observed scene is modeled as a collection of planar patches that are consistent across views, each undergoing a rigid motion that is approximately constant over time. Finding the patches and their motions is cast as minimization of an energy function over the continuous plane and motion parameters and the discrete pixel-to-plane assignment. We show that such a view-consistent multi-frame scheme greatly improves scene flow computation in the presence of occlusions, and increases its robustness against adverse imaging conditions, such as specularities. Our method currently achieves leading performance on the KITTI benchmark, for both flow and stereo.