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.

What problem does this paper attempt to address?

The paper is primarily dedicated to addressing the challenging issues in 3D point cloud registration, particularly by proposing a learning-based matching algorithm to improve existing methods. Specifically, the goals of the paper include: 1. **Improving the matching algorithm**: The authors propose a new learning-based matching algorithm that is jointly optimized with the feature extractor to overcome the limitations of existing handcrafted heuristic algorithms (such as DeSmooth and dual softmax). 2. **Enhancing memory efficiency**: To enable the algorithm to be applied to large-scale 3D point cloud registration tasks, the authors improved an existing two-stage graph matching neural network (i.e., WeaveNet), enhancing its memory efficiency. 3. **Performance improvement**: The authors demonstrate that their proposed matching module can significantly improve the performance of various contemporary feature extractors on rigid and non-rigid, global and partial point cloud registration datasets. The contributions presented in the paper can be summarized as follows: - Proposing a method that combines a learning-based matching algorithm with a feature extractor. - Improving the memory efficiency of an existing two-stage graph matching neural network, making it suitable for 3D point cloud registration. - Achieving state-of-the-art performance in multiple benchmarks. The experimental section validates the effectiveness of the proposed method, particularly in handling non-rigid point cloud registration, achieving significant performance improvements compared to previous best methods. Additionally, analyses on memory consumption, feature uncertainty, and hyperparameter sensitivity were conducted. Finally, the method was also validated on human shape data, demonstrating its effectiveness in both rigid and non-rigid settings.

3D Point Cloud Registration with Learning-based Matching Algorithm

DeepICP: An End-to-End Deep Neural Network for 3D Point Cloud Registration

Point cloud registration method based on deep learning

DeepMatch: Toward Lightweight in Point Cloud Registration

Point Cloud Registration Based on Graph Matching Optimization

Deep Correspondence Matching-Based Robust Point Cloud Registration of Profiled Parts

3D Focusing-and-Matching Network for Multi-Instance Point Cloud Registration

Learning and Matching Multi-View Descriptors for Registration of Point Clouds

An iteration-based interactive attention network for 3D point cloud registration

Learning Sufficient Correlations among Points for 3D Non-rigid Point Cloud Registration

2D3D-MATR: 2D-3D Matching Transformer for Detection-free Registration between Images and Point Clouds

End-to-End 3D Point Cloud Learning for Registration Task Using Virtual Correspondences.

Graph Matching Optimization Network for Point Cloud Registration

3D Point Cloud Registration Method Based on Deep and Shallow Features Combined with Attention Enhancement

A Coarse-to-Fine Algorithm for Matching and Registration in 3D Cross-Source Point Clouds

Robust Point Cloud Registration Framework Based on Deep Graph Matching

Geometry-Aware Enhanced Mutual-Supervised Point Elimination with Overlapping Mask Contrastive Learning for Partitial Point Cloud Registration

End-to-End Learning the Partial Permutation Matrix for Robust 3D Point Cloud Registration

Geometric Transformer for Fast and Robust Point Cloud Registration

A Matching Algorithm for Featureless Sparse Point Cloud Registration.

RRGA-Net: Robust Point Cloud Registration Based on Graph Convolutional Attention