An algorithm for selecting nearest neighbors at a variable scale rather than a fixed search radius in point cloud neighbor search is proposed in this paper. We employ the concepts in differential geometry and divide the point cloud into different clusters according to their surface types. Not only the distance metric but also the clusters' surface type is taken into condition when we search the neighbors of a certain point. This results in a variable scale in nearest neighbor search which can preserve good enough details even using a big scale as well as reduce side effects of noise data caused by using a small scale. The proposed algorithm is tested with the data of Stanford Bunny by simulation. Its effectiveness is confirmed by the experiments.