3D Point Cloud Registration with Learning-based Matching Algorithm

TL;DR

This paper introduces a learning-based matching algorithm for 3D point cloud registration that enhances performance by jointly training feature extractors and improving memory efficiency, significantly boosting state-of-the-art methods.

Contribution

It proposes a novel differential matching module optimized with feature extractors, improving memory efficiency and registration accuracy across various datasets.

Findings

Boosted SOTA method RoITr by +5.4% and +7.2% in NFMR metric.

Achieved +6.1% and +8.5% improvements in IR metric on 4DMatch and 4DLoMatch.

Demonstrated effectiveness on both rigid and non-rigid, whole and partial point cloud datasets.

Abstract

We present a novel differential matching algorithm for 3D point cloud registration. Instead of only optimizing the feature extractor for a matching algorithm, we propose a learning-based matching module optimized to the jointly-trained feature extractor. We focused on edge-wise feature-forwarding architectures, which are memory-consuming but can avoid the over-smoothing effect that GNNs suffer. We improve its memory efficiency to scale it for point cloud registration while investigating the best way of connecting it to the feature extractor. Experimental results show our matching module's significant impact on performance improvement in rigid/non-rigid and whole/partial point cloud registration datasets with multiple contemporary feature extractors. For example, our module boosted the current SOTA method, RoITr, by +5.4%, and +7.2% in the NFMR metric and +6.1% and +8.5% in the IR metric on the 4DMatch and 4DLoMatch datasets, respectively.