Efficient Second-Order Plane Adjustment

Lipu ZhouCVPR 2023 · DOI · arXiv

Abstract

Plane Adjustment (PA) is a fundamental task in 3D reconstruction, analogous to visual Bundle Adjustment, where sensor poses and environmental plane parameters are jointly optimized. While the Levenberg-Marquardt (LM) algorithm is the standard choice for such problems, it often suffers from slow convergence and sensitivity to initial noise. This paper presents an efficient second-order optimization framework for PA. By deriving a closed-form solution for the Hessian matrix that is independent of the number of points on a plane, we enable the practical use of Newton’s method. Our approach achieves significantly faster convergence and superior robustness to noise compared to state-of-the-art LM implementations.

Introduction

In LiDAR-based or RGB-D SLAM, refining the geometry of the environment and the trajectory of the sensor is critical for accurate mapping. Plane Adjustment (PA) minimizes the distance between observed 3D points and the planes they constitute. Traditionally, second-order methods like Newton's method have been deemed computationally prohibitive for large-scale PA due to the complexity of calculating the Hessian matrix. Consequently, most systems rely on first-order approximations like Levenberg-Marquardt. This work challenges that convention by introducing a mathematically optimized Newton's method that is both efficient and robust.

A Plane Adjustment problem illustrating the relationship between sensor poses (X) and environmental planes (π).

Methodology

The proposed method leverages three primary innovations to make second-order optimization feasible for real-time applications:

1. Variable Elimination

The authors observe that for any set of sensor poses, the optimal plane parameters can be determined via a closed-form solution derived from the smallest eigenvalue of a point-scatter matrix. By embedding this solution directly into the cost function, the planes are eliminated as independent variables. The resulting optimization problem depends solely on the sensor poses, significantly reducing the dimensionality of the search space.

2. Closed-Form Hessian Derivation

A key contribution is the derivation of the gradient and Hessian using the Implicit Function Theorem. This allows for the differentiation of the eigenvalue problem without resorting to expensive numerical approximations. This mathematical rigor ensures that the "curvature" of the error surface is captured accurately at every step.

3. Point-Independent Complexity

The computational cost of the Hessian is decoupled from the number of points ($N$) on the planes. By pre-calculating constant matrices (point-scatter matrices) once, the iterative optimization step becomes $O(1)$ with respect to $N$. This allows the algorithm to handle dense point clouds with thousands of points per plane without a performance penalty.

Experimental Evaluation

The algorithm was tested on four datasets captured using a VLP-16 LiDAR, featuring between 472 and 1,606 poses. To evaluate robustness, ground-truth poses were perturbed with Gaussian noise (up to 3° rotation and 0.3m translation).

Visualization of segmented planar features in a LiDAR point cloud used for the optimization experiments.

Results

  • Convergence Speed: The Newton-based method consistently outperformed the Ceres Solver (LM) implementation. In high-noise scenarios, the proposed method converged in a fraction of the iterations required by LM (often 5-10x fewer).
  • Robustness: While LM frequently failed to converge or became trapped in local minima under high noise, the second-order approach remained stable and reached optimal solutions.
  • Efficiency: The point-independent complexity was verified, showing that the per-iteration time remains low even as the density of the environmental scan increases.

Conclusion

This work demonstrates that Newton’s method is not only viable but superior for Plane Adjustment when properly formulated. By eliminating plane variables and deriving a point-independent closed-form Hessian, the authors provide a robust alternative to standard first-order solvers. Future work includes extending this framework to other geometric primitives such as lines and cylinders to support more diverse environments.

Summary updated

BibTeX
@inproceedings{zhou2023efficient,
  author = {Lipu Zhou},
  title = {{Efficient Second-Order Plane Adjustment}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2023},
  doi = {10.1109/cvpr52729.2023.01260},
}
Figure