Yuwei Wu,Weixiao Liu,Sipu Ruan,Gregory S. Chirikjian
Abstract:3D shape abstraction has drawn great interest over the years. Apart from low-level representations such as meshes and voxels, researchers also seek to semantically abstract complex objects with basic geometric primitives. Recent deep learning methods rely heavily on datasets, with limited generality to unseen categories. Furthermore, abstracting an object accurately yet with a small number of primitives still remains a challenge. In this paper, we propose a novel non-parametric Bayesian statistical method to infer an abstraction, consisting of an unknown number of geometric primitives, from a point cloud. We model the generation of points as observations sampled from an infinite mixture of Gaussian Superquadric Taper Models (GSTM). Our approach formulates the abstraction as a clustering problem, in which: 1) each point is assigned to a cluster via the Chinese Restaurant Process (CRP); 2) a primitive representation is optimized for each cluster, and 3) a merging post-process is incorporated to provide a concise representation. We conduct extensive experiments on two datasets. The results indicate that our method outperforms the state-of-the-art in terms of accuracy and is generalizable to various types of objects.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is how to accurately abstract object representations composed of a small number of geometric primitives (especially multi - cones, i.e., superquadrics) from point - cloud data. Specifically, the authors focus on the following two main challenges:
1. **Limitations of existing methods**: Existing deep - learning methods rely on large - scale datasets when abstracting objects, resulting in limited generalization ability and difficulty in handling unseen categories. Moreover, accurately abstracting objects with a small number of primitives remains a challenge.
2. **Improving the accuracy and compactness of abstraction**: How to infer an abstract representation composed of an unknown number of geometric primitives from point - clouds through non - parametric Bayesian inference methods, and ensure that the number of these primitives is as small as possible while maintaining high accuracy.
To solve these problems, the authors propose a new non - parametric Bayesian statistical method, which models the 3D shape abstraction problem as a clustering problem and achieves it through the following steps:
- **Generative model**: Assume that each point in the point - cloud is sampled from the infinite Gaussian super - quadric surface model (GSTM).
- **Clustering process**: Use the Chinese Restaurant Process (CRP) to assign each point to a cluster and optimize the primitive representation of each cluster.
- **Merging post - processing**: Introduce a merging post - processing step to reduce the number of primitives and provide a more concise representation.
### Key formulas
1. **Gaussian super - quadric surface model (GSTM)**:
\[
x\sim\text{GSTM}(\theta,\sigma^{2})
\]
where \(\theta = [\varepsilon_{1},\varepsilon_{2},a_{x},a_{y},a_{z},g,k_{x},k_{y}]\) are the parameters of the super - quadric surface, and \(\sigma^{2}\) is the noise variance.
2. **Mixture model**:
\[
X=\left\{x_{i}\mid x_{i}\sim\sum_{j = 1}^{K}\omega_{j}\text{GSTM}(\theta_{j},\sigma_{j}^{2})\right\}
\]
where \(\sum_{j = 1}^{K}\omega_{j}=1\), and each \(\omega_{j}\) represents the probability that the observed value comes from \((\theta_{j},\sigma_{j}^{2})\).
3. **Dirichlet process (DP)**:
\[
G\sim\text{DP}(G_{0},\alpha)
\]
\[
G=\sum_{j = 1}^{\infty}\omega_{j}\delta_{\Theta_{j}},\quad\Theta_{j}\sim G_{0},\quad\pi\sim\text{GEM}(\alpha)
\]
4. **Sampling distribution**:
\[
p(x\mid\theta,\sigma^{2})\approx\frac{1}{2\sqrt{2\pi\sigma}}\exp\left(-\frac{d^{2}(\theta,x)}{2\sigma^{2}}\right)
\]
where \(d(\theta,x)\) is the distance from point \(x\) to the super - quadric surface.
5. **Merging post - processing**:
For any two clusters, if the merged super - quadric surface can better fit the new point set, then merge these two clusters.
### Experimental results
The authors conducted experiments on the ShapeNet and D - FAUST datasets. The results show that this method is significantly superior to existing methods in terms of accuracy and has better generalization ability. Moreover, through the merging post - processing step, the number of primitives can be reduced while maintaining high accuracy, thereby improving interpretability.
In conclusion, this paper proposes a novel non - parametric Bayesian inference method for abstracting object representations composed of a small number of geometric primitives from point - clouds, solving the existing problems.