A Fast TGV-l 1 RGB-D Flow Estimation

Junha Roh, Hwasup Lim, Sang Chul Ahn

Abstract: We present a novel method for fast and dense 3D scene flow estimation which optimizes consistency and smoothness in both intensity and depth data while considering computing efficiency for the real-world applications. 3D scene flow estimation is an attractive problem with the advent of commodity RGB-D cameras. Naive extensions of recent variational optical flow techniques show promising but limited successes. Due to their primitive priors, solutions from total variation approaches prefer unrealistic constant motion. To overcome these problems and consider the computational efficiency, we adopt an image-guided total generalized variation (ITGV) regularization. As demonstrated with experimental results, the proposed method outperforms both in terms of accuracy and speed compared to the existing variational approaches.