Deep Neighbor Geometric Consistency for 3D Point Cloud Registration*

Due to the difficulty in determining rigid transformations between two partially overlapping point clouds, 3D point cloud registration is still a highly challenging problem in the field of computer vision. Motivated by the recent advent of deep learning-based registration approaches, we propose Deep Neighbor Geometric Consistency, a more adaptable end-to-end alignment network framework for finding corresponding points and predicting rigid transformations. It is based on the following three modules: a deep best-match pointer module for capturing the neighboring geometric details of key points, a pseudo correspondence generation layer for creating correspondence points by learning matching probabilities, and a correspondence weighting layer for removing outliers. We comprehensively evaluate our method’s performance on the ModelNet40 dataset. Experimental results show that our method achieves comparable registration accuracy while being more resilient to initial poses. In particular, DeepNGC can reliably obtain proper pseudo-correspondence consistently across various views and different objects.