Geometry-Aware Enhancement-Based Point Elimination with Overlapping Mask Learning for Partial Point Cloud Registration

Point cloud registration in the low overlap case has been one of the research difficulties in the field. Recent approaches use the Transformer for information interaction to improve the perception between point cloud pairs and the learning of overlapping masks to shield the influence of non-overlapping regions, which significantly improves the accuracy of point cloud registration under low overlapping conditions. However, the former results in low computational efficiency, while the latter leads to less accurate learning results because only feature distance similarity is considered in the mask learning process without taking into account the original geometric structure similarity. In this paper, we propose an improved method. Graph features are first adaptively extracted and fused with PPF geometric features to enhance feature representation. Simultaneous information interaction is performed to enhance the visibility of points in the overlapping region, and then the point features with obvious geometrical structures are retained by the key point extraction module, which improves the efficiency of the subsequent calculations. A Geometry Transformer is then introduced to enhance the perception of geometric similarity between the source and target point clouds after selecting key points. Finally, point features with geometric similarity and distance similarity are further carefully retained by iteratively learning overlapping masks to directly and efficiently regress the transform parameters. Compared with traditional and state-of-the-art deep learning methods, extensive experiments on the ModelNet40 dataset demonstrate that our method exhibits state-of-the-art results in terms of accuracy.