Geometric Transformer for Fast and Robust Point Cloud Registration

Abstract

We study the problem of extracting accurate correspondences for point cloud registration. Recent keypoint-free methods have shown great potential through bypassing the detection of repeatable keypoints which is difficult to do especially in low-overlap scenarios. They seek correspondences over downsampled superpoints, which are then propagated to dense points. Superpoints are matched based on whether their neighboring patches overlap. Such sparse and loose matching requires contextual features capturing the geometric structure of the point clouds. We propose Geometric Transformer, or GeoTransformer for short, to learn geometric feature for robust superpoint matching. It encodes pair-wise distances and triplet-wise angles, making it invariant to rigid transformation and robust in low-overlap cases. The simplistic design attains surprisingly high matching accuracy such that no RANSAC is required in the estimation of alignment transformation, leading to 100 times acceleration. Extensive experiments on rich benchmarks encompassing indoor, outdoor, synthetic, multiway and non-rigid demonstrate the efficacy of GeoTransformer. Notably, our method improves the inlier ratio by 18∼31 percentage points and the registration recall by over 7 points on the challenging 3DLoMatch benchmark. Our code and models are available at https://github.com/qinzheng93/GeoTransformer.

An Efficient and Robust Object-Level Cooperative Perception Framework for Connected and Automated Driving

“…4. Following [21], [32], two metrics are utilized to evaluate the difference between estimated results and ground truth.…”

“…[20] established the problem of finding dense correspondences across semantically similar images as an optimal transport problem. Qin et al [21] applied optimal transport theory on…”

FeatFlow: Learning geometric features for 3D motion estimation

“…These learning-based methods give competitive results, but the performance drops drastically in small overlap scenes. This problem already draws the attention of many researchers, as [12], [13], [15] tried in indoor scenes. Our method only detects keypoints in the estimated overlap, thus avoiding wrong matches between non-overlapping regions.…”

“…Cross attention has demonstrated its effectiveness in interacting information [12], [13] and detecting overlap regions [15] from encoded feature maps. Similar to ImLoveNet [15], we adopt cross attention on two feature maps of the input point clouds to learn relevant information, followed by a classification head to solve the overlap as learning a similarity score.…”

“…Another equally important module in the SLAM system is loop closure, which ensures a globally consistent map. Recent works have made substantial progress in loop closure detection [7]- [10] and point cloud registration [11]- [13]. For loop closure detection, it is common practice to encode the entire point cloud into a global descriptor [7], [8].…”