Abstract: This paper proposes a fully supervised algorithm based on voting and dynamic grouping to extract points of interest on 3D models. In the voting strategy, the neighborhood vertex sequence of vertex is used to generate the probability voting sequence and the confidence sequence to obtain the predicted probability distribution. In dynamic grouping, a grouping strategy is used to extract interest points of interest on the probability distribution and we make a further improvement. Our method is mainly divided into three modules. The first module is a vertex semantic extraction module, which fuses neighborhood and global information of vertices to generate vertex semantic sequences. The second module is the voting network and the confidence network, which maps the vertex semantic sequence into a probability voting sequence and a confidence sequence to generate a probability distribution. The third module is the dynamic grouping module. By setting different probability thresholds to group and extract points of interest from the probability distribution with consideration of points of different probability value levels. Our algorithm was verified on the 3D model data set SHREC2011. The results show that, compared with traditional algorithms, The correct rate of our algorithm to extract points of interest is increased by at least 20%, and the proportion of missed interest points is reduced by at least 18%, Compared with the machine learning algorithm, the correct rate is increased by at least 2.0%, and the missing ratio points of interest is reduced by at least 10%, and the effect of extracting interest points has been significantly improved.