An Improved Discriminative Optimization for 3D Rigid Point Cloud Registration

Jia Wang, Ping Wang, Biao Li, Ruigang Fu, Junzheng Wu

Abstract

Discriminative Optimization (DO) has proven highly effective for 3D point cloud registration by learning to predict transformation parameters through linear regression. However, the original DO's reliance on a simple "front-back" feature descriptor often leads to information loss and optimization failure in sparse or highly misaligned scenarios. This paper presents an improved DO framework that introduces a "triple-binary" descriptor—incorporating front-back, up-down, and clockwise-anticlockwise spatial relationships—and a distribution-based reweighting scheme. Experimental results on the Stanford Bunny and Oxford SensatUrban datasets demonstrate that our method significantly enhances registration accuracy and robustness compared to state-of-the-art algorithms.

Introduction

Point cloud registration is a fundamental task in computer vision, requiring the precise alignment of a "scene" point cloud to a "model" reference. Traditional methods like Iterative Closest Point (ICP) often struggle with local minima and noise. Discriminative Optimization (DO) addresses this by learning a sequence of linear maps that guide the registration process. Despite its promise, the original DO's feature descriptor was insufficient for capturing complex rotational misalignments. We propose an enhanced descriptor and a refined optimization strategy to overcome these limitations.

Methodology: The Triple-Binary Descriptor

The primary contribution of this work is the expansion of the DO feature space. Instead of a binary "front-back" classification, we propose a triple-binary system that captures three-dimensional spatial context relative to model points.

The triple-binary descriptor captures spatial relationships across three dimensions: (a) front-back, (b) up-down, and (c) clockwise-anticlockwise.

By encoding elevation (up-down) and azimuth (clockwise-anticlockwise) angles, the descriptor provides the regressor with a much richer representation of rotational misalignment. Furthermore, we implement a distribution-based reweighting of the histogram bins, allowing the model to prioritize the most informative regions of the point cloud.

To ensure mathematical rigor, the optimization is parameterized using Lie groups ($SE(3)$) and Lie algebras ($se(3)$). This allows for a unified linear representation of 3D rotations and translations. We also provide a formal proof (Theorem 1) demonstrating that our sequential update rule ensures a strict decrease in training error, guaranteeing convergence toward the optimal alignment.

Experimental Setup and Datasets

The proposed algorithm was evaluated using both synthetic and real-world data:

  1. Stanford Bunny: A standard synthetic benchmark used to evaluate performance under various perturbations, including noise, outliers, and varying point densities.
  2. Oxford SensatUrban: A large-scale urban LIDAR dataset. We utilized the Birmingham subset, which presents significant challenges due to unclosed surfaces and complex environmental noise.
The Stanford Bunny dataset serves as a key benchmark for testing registration robustness against synthetic noise and outliers.

Results and Discussion

Our "Improved DO" was compared against six state-of-the-art algorithms, including ICP, CPD, and BCPD. Key findings include:

  • Enhanced Accuracy: On the Stanford Bunny dataset, the Improved DO achieved a Point Accuracy of 0.990 under noise, a significant improvement over the original DO's 0.863.
  • Real-World Performance: In the SensatUrban tests, our method achieved 0.993 accuracy, outperforming BCPD (0.743) and the original DO (0.777).
  • Robustness to Sparsity: The enriched descriptor effectively prevents the "all-zero" vector problem, maintaining stability even when the initial scene-model offset is large.
  • Convergence: The algorithm demonstrates faster convergence and higher precision than its predecessor, requiring fewer iterations to achieve optimal alignment.

Conclusion

By enriching the feature descriptor and optimizing the learning framework, we have significantly improved the performance of Discriminative Optimization for 3D rigid point cloud registration. The "triple-binary" approach effectively captures complex spatial relationships, leading to superior accuracy and robustness in both synthetic and real-world urban environments. Future work will focus on extending these principles to non-rigid registration and handling extreme initial translations.

Summary updated

Abstract

The Discriminative Optimization (DO) algorithm has been proved much successful in 3D point cloud registration. In the original DO, the feature (descriptor) of two point cloud was defined as a histogram, and the element of histogram indicates the weights of scene points in "front" or "back" side of a model point. In this paper, we extended the histogram which indicate the sides from "front-back" to "front-back", "up-down", and "clockwise-anticlockwise". In addition, we reweighted the extended histogram according to the model points' distribution. We evaluated the proposed Improved DO on the Stanford Bunny and Oxford SensatUrban dataset, and compared it with six classical State-Of-The-Art point cloud registration algorithms. The experimental result demonstrates our algorithm achieves comparable performance in point registration accuracy and root-mean-sqart-error.
BibTeX
@misc{wang2021improved,
  author = {Jia Wang and Ping Wang and Biao Li and Ruigang Fu and Junzheng Wu},
  title = {{An Improved Discriminative Optimization for 3D Rigid Point Cloud Registration}},
  year = {2021},
  eprint = {2104.08854},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2104.08854},
}
Figure