Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Strong Baseline for Point Cloud Registration via Direct Superpoints Matching

Aniket Gupta, Yiming Xie, Hanumant Singh, Huaizu Jiang

TL;DR

This work targets robust point cloud registration by directly matching downsampled superpoints across two clouds. It eliminates reliance on RANSAC or predicting non-existent point correspondences by computing a global similarity matrix via Softmax-based correlations and weighing inliers in a differentiable, end-to-end framework that estimates the $SE(3)$ transformation with a weighted Kabsch-Umeyama solver. The KPConv-based superpoint extraction, Transformer-based feature enhancement, and correlation-weighted pose estimation together yield competitive or superior results on ModelNet, 3DMatch, and KITTI, while avoiding tedious postprocessing. The findings emphasize that a strong, well-calibrated matching strategy can significantly boost registration accuracy and efficiency across diverse environments, including outdoor LiDAR data.

Abstract

Deep neural networks endow the downsampled superpoints with highly discriminative feature representations. Previous dominant point cloud registration approaches match these feature representations as the first step, e.g., using the Sinkhorn algorithm. A RANSAC-like method is then usually adopted as a post-processing refinement to filter the outliers. Other dominant method is to directly predict the superpoint matchings using learned MLP layers. Both of them have drawbacks: RANSAC-based methods are computationally intensive and prediction-based methods suffer from outputing non-existing points in the point cloud. In this paper, we propose a straightforward and effective baseline to find correspondences of superpoints in a global matching manner. We employ the normalized matching scores as weights for each correspondence, allowing us to reject the outliers and further weigh the rest inliers when fitting the transformation matrix without relying on the cumbersome RANSAC. Moreover, the entire model can be trained in an end-to-end fashion, leading to better accuracy. Our simple yet effective baseline shows comparable or even better results than state-of-the-art methods on three datasets including ModelNet, 3DMatch, and KITTI. We do not advocate our approach to be \emph{the} solution for point cloud registration but use the results to emphasize the role of matching strategy for point cloud registration. The code and models are available at https://github.com/neu-vi/Superpoints_Registration.

A Strong Baseline for Point Cloud Registration via Direct Superpoints Matching

TL;DR

This work targets robust point cloud registration by directly matching downsampled superpoints across two clouds. It eliminates reliance on RANSAC or predicting non-existent point correspondences by computing a global similarity matrix via Softmax-based correlations and weighing inliers in a differentiable, end-to-end framework that estimates the transformation with a weighted Kabsch-Umeyama solver. The KPConv-based superpoint extraction, Transformer-based feature enhancement, and correlation-weighted pose estimation together yield competitive or superior results on ModelNet, 3DMatch, and KITTI, while avoiding tedious postprocessing. The findings emphasize that a strong, well-calibrated matching strategy can significantly boost registration accuracy and efficiency across diverse environments, including outdoor LiDAR data.

Abstract

Deep neural networks endow the downsampled superpoints with highly discriminative feature representations. Previous dominant point cloud registration approaches match these feature representations as the first step, e.g., using the Sinkhorn algorithm. A RANSAC-like method is then usually adopted as a post-processing refinement to filter the outliers. Other dominant method is to directly predict the superpoint matchings using learned MLP layers. Both of them have drawbacks: RANSAC-based methods are computationally intensive and prediction-based methods suffer from outputing non-existing points in the point cloud. In this paper, we propose a straightforward and effective baseline to find correspondences of superpoints in a global matching manner. We employ the normalized matching scores as weights for each correspondence, allowing us to reject the outliers and further weigh the rest inliers when fitting the transformation matrix without relying on the cumbersome RANSAC. Moreover, the entire model can be trained in an end-to-end fashion, leading to better accuracy. Our simple yet effective baseline shows comparable or even better results than state-of-the-art methods on three datasets including ModelNet, 3DMatch, and KITTI. We do not advocate our approach to be \emph{the} solution for point cloud registration but use the results to emphasize the role of matching strategy for point cloud registration. The code and models are available at https://github.com/neu-vi/Superpoints_Registration.
Paper Structure (13 sections, 6 equations, 4 figures, 7 tables)

This paper contains 13 sections, 6 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Superpoints Feature Extraction and Enhancement
  5. Superpoint Matching for SE(3) Transformation Estimation
  6. Loss Functions
  7. Experiments
  8. Implementation details
  9. ModelNet and ModelLoNet Benchmarks
  10. 3DMatch and 3DLoMatch Benchmarks
  11. KITTI Benchmark
  12. Ablation Studies
  13. Conclusion

Figures (4)

  • Figure 1: Our approach directly matches the superpoints between the input point clouds to register them by estimating the $SE(3)$ transformation matrix. The correlation weights obtained from matching are used to filter out incorrect correspondences (outliers) and further weigh the rest inliers for the transformation matrix estimation. The source cloud is shown in blue, with superpoints highlighted in red. Similarly, the target point cloud is displayed in yellow, with matching superpoints shown in blue. The intensity of the red and blue colors represents the correspondence weights. Notice how most keypoints are distributed around the three-way junction and exhibit high correspondence weights. (Best viewed in color.)
  • Figure 2: Model Architecture: The KPConv backbone downsamples the input point cloud and generates superpoints and their feature vectors. These superpoint features are then conditioned on the other point cloud in the feature enhancement block. Lastly, superpoint features can be directly matched using Global Softmax to estimate SE(3) transformation using the corrleation weights in a robust manner. The intensity of the red and blue colors of the superpoints after the softmax correlation step represents the correspondence weights. (Best viewed in color.)
  • Figure 3: Correlation weight distribution on 3DMatch. Here, many points exhibit moderate correlations, suggesting that a considerable portion of sample points are reasonably good matches. In contrast, 3DLoMatch has fewer points that exhibit strong correlations due to the low data overlap.
  • Figure 4: Qualitative Results on 3DMatch. The source cloud is shown in blue with the red color representing the matched superpoints. Similarily target cloud is shown in yellow color with matching superpoints in blue. (Best viewed in color.)