Parallel computation of high-accuracy scene flow based on multiview stereo

场景流是目标物体在真实空间中运动的三维矢量场表示，它可以从不同视点拍摄的一组图像序列中估计出来。近年来，在CG、VR、运动分析等各种应用中，都需要快速准确的场景流估计方法，我们的研究就是针对这一问题进行的。作为进展的一个标志，我们开发了一种基于体素的多视点立体场景流估计方法，并在PC集群机器上实现了这种场景流估计的并行方法。此外，我们还通过几个应用程序的实验证明了我们所提出的方法的有效性。

Scene flow is 3D vector field representation of target object's motion in a real space, which can be estimated from a set of image sequences taken at different viewpoints. Recently, high-speed methods for estimating accurate scene flow are needed in widely various applications such as CG, VR, motion analysis, etc. Our research addresses this issue. As an indication of our progress, we have developed a scene flow estimation method with voxel based multiview stereo and a parallel method for performing this scene flow estimation on a PC cluster machine. Furthermore, we demonstrate the effectiveness of our developed methods through the experiments on several applications.