Abstract: Aiming at the complicated post-processing and poor generalization ability of correspondence calculation of non-rigid 3D point cloud shapes, a method that employs a mixed attention mechanism and unsupervised learning to calculate correspondence is proposed in this paper. First, the point pair feature improves the EdgeConv so that the extracted features can contain more similar pose information between points. Then, a mixed attention similarity refinement module is constructed by calculating cosine similarity, and the similar parts of features between models encode as a similarity matrix. Finally, the corresponding model is directly reconstructed in both directions using the similarity matrix and the coordinate information to compute the final correspondence. The qualitative and quantitative experimental results on SURREAL, SHREC’19, SMAL, and TOSCA datasets show that the proposed algorithm outperforms the CorrNet3D algorithm. Specifically, the average error in measuring correspondence error using the Euclidean distance between the original and reconstructed shapes is reduced by 0.19 and 5.0 on the SHREC’19 and TOSCA datasets, respectively. The correspondence accuracy is also improved by 9.26% and 20.41% when the tolerance error is 10%. Furthermore, the proposed algorithm exhibits good generalization ability across different datasets.