Motion segmentation for dynamic scene is a fundamental problem in computer vision due to its well-known chicken-and-egg character. The key issue is to estimate both numbers and parameters of motions simultaneously. Different from global clustering method and random sampling scheme, in this paper, we propose a divided-and-conquer algorithm to solve the motion segmentation problem. A guided selection is used to choose the most creditable hypothetical motion as a candidate seed and then make it grow larger. Compared to previous works such as expectation maximization and factorization approaches, there is no need for any pre-knowledge of the number of motions. To global non-parametric clustering method, it is fast because each time we only do cluster process in a partitioned sub-set. Experiments have shown that the proposal method can give a satisfying result for motion segmentation.