总结一下The main solution for large-scale point clouds registration is to first obtaina set of matched3D keypoint pairs and then accomplish the point cloud registration task based on these matched keypoin

This study focuses on addressing the 3D keypoints detection problem for large-scale point clouds registration. The commonly used voxel-grid-based downsampling method leads to a low inlier ratio. The study proposes four 3D keypoints detection methods based on the joint keypoint detection and description learning framework D3Feat, using deep learning. The Multi-layer Perceptron (MLP) based method achieves the best inlier ratios and state-of-the-art registration performance on both indoor and outdoor large-scale point clouds datasets. This study highlights the importance of considering 3D keypoints detection in point clouds registration tasks and proposes a promising approach using deep learning.