Deep Learning for 3D Keypoint Detection in Large-Scale Point Cloud Registration

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.