3D mesh segmentation has been widely used in fields such as Computer Vision. Existing algorithms have the problems of over- and under-merging when performing 3D mesh segmentation. In this paper, we propose a robust and efficient segmentation algorithm using a so-called two-stage merging strategy. We first define a flatness measure for mesh faces, and partition the given mesh model into different patches based on the flatness measure. In order to handle the over- and under-merging problems, we first perform a local merging operation and then a global merging operation. Experiments show that our algorithm is effective, and works well even for models without obvious part boundaries at their joints.