Robust and Accurate Superquadric Recovery: a Probabilistic Approach

Weixiao Liu, Yuwei Wu, Sipu Ruan, Gregory S. ChirikjianCVPR 2022 · OpenReview

Introduction

In the fields of computer vision and robotics, representing complex 3D environments using simple geometric primitives is a fundamental challenge. Superquadrics offer a versatile solution, capable of modeling a wide range of shapes—from ellipsoids to sharp-edged boxes—with a minimal set of parameters. However, traditional recovery methods based on Least-Squares (LSQ) optimization often struggle with local optima, sensor noise, and numerical instability, particularly when modeling objects with sharp edges.

To address these limitations, researchers have introduced a novel probabilistic framework called EMS (Expectation, Maximization, and Switching). This approach redefines shape recovery as a Maximum Likelihood Estimation (MLE) problem, significantly improving robustness and accuracy in real-world scenarios.

The EMS Framework: A Probabilistic Solution

The core innovation of this work is the EMS algorithm, which utilizes a Gaussian-Uniform Mixture (GUM) model to distinguish between valid shape data and noise. The algorithm operates in three distinct phases:

  1. Expectation (E-step): The model calculates the probability of each point in a 3D cloud being an "inlier" (part of the object) or an "outlier" (noise). This explicit modeling allows the system to naturally ignore clutter.
  2. Maximization (M-step): Superquadric parameters are updated to maximize the likelihood of the inlier points. By employing a trust-region reflective algorithm, the method achieves superior stability for sharp-edged shapes compared to standard solvers.
  3. Switching (S-step): This step addresses the "algebraic ambiguity" of superquadrics, where different parameters can represent nearly identical shapes. The S-step allows the algorithm to "jump" between these similar parameter sets to escape local optima and find the global best fit.
The EMS algorithm uses a switching mechanism to escape local optima by jumping between geometrically similar parameter sets.

Hierarchical and Multi-Part Recovery

Beyond single-shape fitting, the framework extends to Multi-superquadric Recovery. Using a hierarchical decomposition strategy, complex objects—such as household bottles or toy animals—are broken down into a collection of simpler superquadric parts. This allows for a more nuanced representation of intricate geometries while maintaining the simplicity of the underlying primitives.

Performance and Robustness

The EMS method was rigorously evaluated against synthetic datasets and high-fidelity scans (KIT ObjectModel and BigBIRD). The results demonstrate a significant leap in performance:

  • Accuracy: On the KIT dataset, EMS achieved an average error of 0.584 mm, outperforming previous state-of-the-art methods like the Numerical Stable (0.638 mm) and Robust-fitting (0.748 mm) algorithms.
  • Efficiency: Despite its probabilistic nature, the method is highly efficient, processing objects in the BigBIRD dataset in just 78.9 ms on average.
  • Outlier Resilience: As shown in comparative tests, EMS maintains high accuracy even when 40% of the data consists of outliers, whereas deterministic baseline methods often fail to recover the correct shape.
Comparison of EMS against baseline methods across varying outlier ratios, demonstrating superior accuracy and stability.

Conclusion and Future Directions

The EMS framework provides a robust, accurate, and efficient path for superquadric recovery from noisy 3D data. By embracing parameter ambiguity through the Switching step and utilizing a probabilistic model for noise, it overcomes the primary hurdles of traditional LSQ methods. Future work aims to expand the shape vocabulary further by incorporating non-convex primitives and primitive subtractions to model even more complex real-world objects.

Summary updated

Abstract

Interpreting objects with basic geometric primitives has long been studied in computer vision. Among geometric primitives, superquadrics are well known for their ability to represent a wide range of shapes with few parameters. However, as the first and foremost step, recovering superquadrics accurately and robustly from 3D data still remains challenging. The existing methods are subject to local optima and sensitive to noise and outliers in real-world scenarios, resulting in frequent failure in capturing geometric shapes. In this paper, we propose the first probabilistic method to recover superquadrics from point clouds. Our method builds a Gaussian-uniform mixture model (GUM) on the parametric surface of a superquadric, which explicitly models the generation of outliers and noise. The superquadric recovery is formulated as a Maximum Likelihood Estimation (MLE) problem. We propose an algorithm, Expectation, Maximization, and Switching (EMS), to solve this problem, where: (1) outliers are predicted from the posterior perspective; (2) the superquadric parameter is optimized by the trust-region reflective algorithm; and (3) local optima are avoided by globally searching and switching among parameters encoding similar superquadrics. We show that our method can be extended to the multi-superquadrics recovery for complex objects. The proposed method outperforms the state-of-the-art in terms of accuracy, efficiency, and robustness on both synthetic and real-world datasets. The code is at http://github.com/bmlklwx/EMS-superquadric_fitting.git.
BibTeX
@inproceedings{liu2022robust,
  author = {Weixiao Liu and Yuwei Wu and Sipu Ruan and Gregory S. Chirikjian},
  title = {{Robust and Accurate Superquadric Recovery: a Probabilistic Approach}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2022},
  url = {https://openreview.net/forum?id=FGD0tPt_uq},
}
Figure