This paper addresses the problem of simultaneously estimation of 3D structure and motion of moving surfaces from multi-view sequences. We model the surface locally as a 9-parameter spatio-temporal plane and a region growing mechanism is used to guarantee proper initial values are provided for parameter estimation. The algorithm starts with feature points that have been matched across views and time, and dense structure and motion then grow from these points. As demonstrated in the experiment, our method is able to handle large motion, topology change and frequent self-occlusion and the resulted 3D structure and motion are dense and accurate which is particularly useful for applications that require precise quantitative analysis of 3D motion.