Abstract:Accurate and efficient point cloud registration is a challenge because the noise and a large number of points impact the correspondence search. This challenge is still a remaining research problem since most of the existing methods rely on correspondence search. To solve this challenge, we propose a new data-driven registration algorithm by investigating deep generative neural networks to point cloud registration. Given two point clouds, the motivation is to generate the aligned point clouds directly, which is very useful in many applications like 3D matching and search. We design an end-to-end generative neural network for aligned point clouds generation to achieve this motivation, containing three novel components. Firstly, a point multi-perception layer (MLP) mixer (PointMixer) network is proposed to efficiently maintain both the global and local structure information at multiple levels from the self point clouds. Secondly, a feature interaction module is proposed to fuse information from cross point clouds. Thirdly, a parallel and differential sample consensus method is proposed to calculate the transformation matrix of the input point clouds based on the generated registration results. The proposed generative neural network is trained in a GAN framework by maintaining the data distribution and structure similarity. The experiments on both ModelNet40 and 7Scene datasets demonstrate that the proposed algorithm achieves state-of-the-art accuracy and efficiency. Notably, our method reduces $2\times$ in registration error (CD) and $12\times$ running time compared to the state-of-the-art correspondence-based algorithm.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the accuracy and efficiency challenges in point cloud registration. Specifically, existing point cloud registration methods usually rely on corresponding point search, which is not effective when dealing with noise and a large number of points. The paper proposes a new data - driven registration algorithm. By studying deep generative neural networks to directly generate aligned point clouds, it bypasses the step of corresponding point search required in traditional methods. This method not only improves the registration accuracy but also significantly reduces the running time, especially in the presence of a large amount of noise. ### Main Contributions 1. **Directly Generate Registration Results Using Generative Neural Networks for the First Time**: This is the first method to directly generate point cloud registration results using generative neural networks. 2. **Propose the PointMixer Module**: This module can capture the global and local structural information of point clouds and maintain this information at multiple levels. 3. **Feature Interaction Module**: This module is used to fuse information from different point clouds. 4. **Parallel and Differential Sample Consistency (PDSAC) Method**: This method can estimate the transformation matrix in an end - to - end learning framework. 5. **Multiple Loss Functions**: These loss functions are used to train the generative network to maintain the structural similarity of the input point clouds and the implicit data distribution. ### Technical Details - **PointMixer**: Combines the channel mixing and token mixing operations of the multi - layer perceptron (MLP) to maintain the global and local structural information of point clouds. - **Feature Interaction Module**: Fuses the structural information between different point clouds through the concatenation of max - pooling and feature differences. - **PDSAC**: Estimates the transformation matrix using the correspondence between the generated point clouds and the original point clouds. - **Loss Functions**: - **Absolute Loss**: Uses the Earth Mover's Distance (EMD) to measure the difference between the generated point cloud and the target point cloud. - **Relative Loss**: Uses the Mean Absolute Error (MAE) to measure the structural similarity between the original point cloud and its generated point cloud. - **Cyclic Consistency Loss**: Ensures that the generated point cloud can return to its original state after the inverse transformation. - **Adversarial Loss**: Uses the Generative Adversarial Network (GAN) to maintain the consistency of the data distribution. - **Transformation Loss**: Calculates the difference between the estimated transformation matrix and the real transformation matrix. ### Experimental Results - **ModelNet40 Dataset**: On this dataset, the proposed GenReg method outperforms existing methods in both registration accuracy and speed. Specifically, the Chamfer Distance (CD) is reduced by a factor of 2, and the running time is reduced by a factor of 12. - **7Scenes Dataset**: On real - scene data, GenReg also performs well, especially when dealing with noise and partial overlap problems. - **Robustness Test**: After adding Gaussian noise, GenReg still maintains a high registration quality and transformation estimation accuracy. - **Generalization Ability**: On unseen CAD models, GenReg still performs well, demonstrating its strong generalization ability. ### Conclusion The paper proposes a point cloud registration method GenReg based on generative neural networks. This method avoids the complexity and instability of corresponding point search in traditional methods by directly generating aligned point clouds, thus achieving a significant improvement in registration accuracy and efficiency.

GenReg: Deep Generative Method for Fast Point Cloud Registration

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

PointRegGPT: Boosting 3D Point Cloud Registration using Generative Point-Cloud Pairs for Training

DOPNet: Achieving Accurate and Efficient Point Cloud Registration Based on Deep Learning and Multi-Level Features

Deep learning-based point cloud registration: a comprehensive investigation

HRegNet: A Hierarchical Network for Efficient and Accurate Outdoor LiDAR Point Cloud Registration

Two-Stage Point Cloud Registration Framework Based on Graph Neural Network and Attention

Point Cloud Registration Method Based on Geometric Constraint and Transformation Evaluation

GERA: Geometric Embedding for Efficient Point Registration Analysis

GaussReg: Fast 3D Registration with Gaussian Splatting

PACNet: A High-precision Point Cloud Registration Network Based on Deep Learning

Robust Point Cloud Registration Framework Based on Deep Graph Matching

GO: A two-step generative optimization method for point cloud registration

HRegNet: A Hierarchical Network for Large-scale Outdoor LiDAR Point Cloud Registration

Point cloud registration for LiDAR and photogrammetric data: A critical synthesis and performance analysis on classic and deep learning algorithms

PDC-Net: Robust point cloud registration using deep cyclic neural network combined with PCA

VPRNet: Virtual Points Registration Network for Partial-to-Partial Point Cloud Registration

Point Cloud Registration using Representative Overlapping Points

DeepGMR: Learning Latent Gaussian Mixture Models for Registration

Sampling Network Guided Cross-Entropy Method for Unsupervised Point Cloud Registration