Abstract: Local entropy of data points changing sharply over neighborhood is introduced as a detection criterion to classify points as a feature, where the local surface curvature changes greatly. Repeatability rate is introduced as well to reflect the frequency that a sample point is detected as a feature point during its verification at different sizes of local windows. Experiments show that such a multi-scale feature point detection approach can improve the reliability of the algorithm. Furthermore, non-uniformly sampled point cloud can be dealt with.