Construction of Feature Tensor Descriptor and Self-Similarity Analysis for 3D Point Cloud Models

Hailong Hu, Li Zhong, Shengwei Qin, Lizhuang MaJournal of Computer-Aided Design & Computer Graphics 2021 · DOI

Abstract

Local self-similarity is a fundamental challenge in 3D shape analysis. This paper proposes a self-similarity analysis method based on a tensor fusion feature descriptor. By approximately calculating the Shape Diameter Function (SDF) for point clouds and combining it with Shape Index (SI) and Gauss Curvature (GS) into a 3D feature tensor, the method creates a robust local shape descriptor. Experimental results on benchmark models demonstrate that this multi-feature fusion significantly improves the recognition accuracy of similar sub-blocks compared to single-feature approaches.

Introduction

Detecting "local self-similarity"—repeating geometric structures within a single model—is crucial for 3D model compression, retrieval, and alignment. Traditional methods often rely on single geometric features like Gaussian curvature, which may lack the descriptive power to distinguish complex shapes. Furthermore, many algorithms are optimized for meshes, making them difficult to apply to raw, unstructured point cloud data. This research addresses these gaps by introducing a tensor-based descriptor that fuses multiple geometric properties.

Methodology

The proposed pipeline for self-similarity analysis involves four key stages:

1. Point Cloud SDF Approximation

The Shape Diameter Function (SDF) measures the local thickness of a model. Since raw point clouds lack surface connectivity, the authors use Axis-Aligned Bounding Boxes (AABB) to identify relevant facets and antipodal points, allowing for a direct estimation of thickness from the point distribution.

2. Spectral Over-segmentation

To analyze local units, the model is partitioned into approximately 100 sub-blocks using spectral clustering. This segmentation provides the basis for comparing different regions of the model.

3. Feature Tensor Construction

For each point within a sub-block, three distinct geometric features are extracted:

  • Gauss Curvature (GS): Captures local surface bending.
  • Shape Index (SI): Categorizes local topography (e.g., ridges vs. valleys).
  • SDF: Represents the global thickness of the part.

These features are organized into a 3D tensor for each point and its $K$-nearest neighbors. This structure preserves the spatial relationships between a point and its neighborhood across multiple feature types.

Visualization of Shape Index (SI) and Shape Diameter Function (SDF) features on a bull model.

4. Similarity Mapping and Analysis

The Frobenius norm of the feature tensor is used to create a mapping that serves as the final shape descriptor. Similarity between sub-blocks is then calculated by comparing these descriptors, enabling the identification of self-similar parts.

Results and Discussion

The method was evaluated using the Princeton Shape Benchmark (PSB). Quantitative analysis on the Armadillo model showed a relative error in point count ($E_p$) of 0.044, a significant improvement over existing methods (which ranged from 0.098 to 1.219).

Results of self-similarity analysis showing matched sub-blocks on a bull model.

Qualitatively, the algorithm successfully identified symmetrical components, such as the wings of a bird or the legs of a cow, even in the presence of minor noise. The tensor fusion approach proved more resilient than single-feature methods, as the combination of curvature, topography, and thickness creates a unique "fingerprint" for each sub-block.

Conclusion

By fusing multiple geometric features into a 3D tensor, this method provides a powerful tool for point cloud shape analysis. It overcomes the limitations of single-feature descriptors and mesh-dependency, offering high accuracy and computational efficiency for detecting self-similarity in complex 3D models.

Reference: Hu, H., Li, Z., Qin, S., & Ma, L. (2021). Construction of Feature Tensor Descriptor and Self-Similarity Analysis for 3D Point Cloud Models. Journal of Computer-Aided Design & Computer Graphics, 33(4), 590-600.

Summary updated

Abstract

<p indent=0mm>Local self-similarity of 3D model is a fundamental problem in the shape analysis. The construction of a local shape descriptor is very important to the final result of self-similarity analysis. To solve this problem, a self-similarity analysis method based on the tensor fusion feature descriptor is proposed. Firstly, the shape diameter function (SDF) of a point cloud model is approximately calculated by using relevant facets and antipodal points. Then, spectral clustering is used to segment the model into sub-blocks, and the three-dimensional feature tensor is constructed from the SDF, shape index (SI) and Gauss curvature (GS) matrix of <italic>K</italic>NN neighborhood points. Finally, the shape descriptor is obtained by constructing the mapping with the tensor norm, and then the similarity measure is defined and the self-similarity between the sub-blocks of the model is analyzed. Several state-of-the-art methods (including partial matching and saliency detection) are tested. In terms of not only the visual effect, but also the similarity measure and the relative errors, the results show that this method can effectively describe the shape and improves the recognition accuracy of similar sub-blocks of a point cloud model.
BibTeX
@article{hu2021construction,
  author = {Hailong Hu and Li Zhong and Shengwei Qin and Lizhuang Ma},
  title = {{Construction of Feature Tensor Descriptor and Self-Similarity Analysis for 3D Point Cloud Models}},
  journal = {Journal of Computer-Aided Design \& Computer Graphics},
  year = {2021},
  doi = {10.3724/sp.j.1089.2021.18542},
}
Figure