MFINet: a multi-scale feature interaction network for point cloud registration

Point cloud registration is widely applied in computer vision field. Previous learning-based registration methods focus on extracting global information from the input and ignore local neighborhood information, which makes it difficult to further improve the registration accuracy. In this work, we present MFINet, a multi-scale feature interaction network that can extract multi-scale features and adaptively fuse these features. To achieve this, we use the K-nearest neighbors algorithm to search for the neighborhood of each point in the input, extract local features at different scales in parallel, and perform feature interactions at the feature extractor to enhance the information correlation between the inputs. Next, we propose a multi-scale feature fusion module that learns appropriate weights for each feature extractor branch and fuses these multi-scale features by weighted combination to enhance the representation ability of features. Finally, a dual-branch structure is used to predict the rotation quaternion and translation vector to mitigate the influence of the disparity in solution space between rotation and translation on the registration performance. Experiments on the ModelNet40 dataset show that the MFINet outperforms previous methods in terms of registration accuracy and robustness against noise, and also exhibits a stable generalization capability on the Stanford 3D Scan dataset. Code is available at https://github.com/daqi01/MFINet-master.