Search

• FRN layer can improve performance and reduce the dependence on batch size. • ECA DGCNN can enhance the feature learning ability of the model on the graph structure data. • FEPoinTr achieves superior performance on multiple data sets. • The performance loss of FEPoinTr in small batches is significantly less than that of other models. The point cloud data generated by LiDAR sensors has rich spatial geometry information, which is crucial for many fields. However, due to the limitations of the sensor, the point cloud with incomplete geometric information is usually obtained. To solve this problem, we propose FEPoinTr, which improves performance and reduces batch size dependency by introducing an FRN layer to normalize. In addition, the ECA DGCNN network is used to enhance the feature learning ability of the model on the graph structure data. We verified the advantages of FEPoinTr on ShapeNet-55, PCN, and KITTI datasets, achieving 0.90CD on ShapeNet-55, 7.52CD on PCN, and 0.504MMD on real point cloud dataset KITTI. Both index results and visualization results are better than other models. It is worth noting that when the batch size is reduced to 1, the CD index on the ShapeNet-55 is 1.13, the F-Score index is 0.503, and the performance loss at small batches is significantly less than that of other models. [ABSTRACT FROM AUTHOR]