CP-Census: A Novel Model for Dense Variational Scene Flow from RGB-D Data.

摘要: We present a novel method for dense variational scene flow estimation based multiscale Ternary Census Transform in combination with patchwise Closest Points depth data term. On the one hand, intensity term is capable of handling illumination changes, low texture and noise. other search increases robustness structured regions. Further, we utilize higher order regularization which weighted directed according to input by an anisotropic diffusion tensor. This allows calculate accurate field supports smooth as well non-rigid movements while preserving boundaries. The numerical algorithm solved on primal-dual formulation efficiently parallelized run at high frame rates. In extensive qualitative quantitative evaluation show that this calculation outperforms existing approaches. applicable any sensor delivering such Microsoft Kinect or Intel Gesture Camera.