[Registration] Rigid point cloud registration

Reference

REGTR: End-to-end Point Cloud Correspondences with Transformers

Rigid point cloud registration

Rigid point cloud registration refers to the problem of finding the optimal rotation and translation parameters that align two point clouds.

A common solution to point cloud registration follows the following pipeline:

1. detect salient keypoints
2. compute feature descriptors for these keypoints
3. obtain putative correspondences via nearest neighbor matching
4. estimate the rigid transformation, typically in a robust fashion using RANSAC.

In recent years, researchers have applied learning to point cloud registration.

Many of these works focus on learning the feature descriptors [14,15,54] and sometimes also the keypoint detection [2, 20, 49].

The final two steps generally remain unchanged and these approaches still require nearest neighbor matching and RANSAC to obtain the final transformation.

These algorithms do not take the post-processing into account during training, and their performance can be sensitive to the post-processing choices to pick out the correct correspondences, e.g. number of sampled interest points or distance threshold in RANSAC.