Point cloud completion network for 3D shapes with morphologically diverse structures

Chun-Jing Si,Zhi-Ben Yin,Zhen-Qi Fan,Fu-Yong Liu,Rong Niu,Na Yao,Shi-Quan Shen,Ming-Deng Shi,Ya-Jun Xi

DOI: https://doi.org/10.1007/s40747-023-01325-8

IF: 6.7

2024-02-02

Complex & Intelligent Systems

Abstract:Abstract Point cloud completion is a challenging task that involves predicting missing parts in incomplete 3D shapes. While existing strategies have shown effectiveness on point cloud datasets with regular shapes and continuous surfaces, they struggled to manage the morphologically diverse structures commonly encountered in real-world scenarios. This research proposed a new point cloud completion method, called SegCompletion, to derive complete 3D geometries from a partial shape with different structures and discontinuous surfaces. To achieve this, morphological segmentation was introduced before point cloud completion by deep hierarchical feature learning on point sets, and thus, the complex morphological structure was segmented into regular shapes and continuous surfaces. Additionally, each instance of a point cloud that belonged to the same type of feature could also be effectively identified using HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise). Furthermore, the multiscale generative network achieved sophisticated patching of missing point clouds under the same geometric feature based on feature points. To compensate for the variance in the mean distances between the centers of the patches and their closest neighbors, a simple yet effective uniform loss was utilized. A number of experiments on ShapeNet and Pheno4D datasets have shown the performance of SegCompletion on public datasets. Moreover, the contribution of SegCompletion to our dataset (Cotton3D) was discussed. The experimental results demonstrated that SegCompletion performed better than existing methods reported in the literature.

computer science, artificial intelligence

What problem does this paper attempt to address?

### The Problem Addressed by the Paper

This paper aims to address a challenging problem in point cloud completion, which is how to generate a complete 3D geometry from partial 3D shapes with complex morphological structures and discontinuous surfaces. Existing point cloud completion methods perform well on datasets with regular shapes and continuous surfaces, but in the real world, many objects have diverse morphological structures and discontinuous surfaces, making it difficult for existing methods to cope.

Specifically, the paper proposes a new point cloud completion method called SegCompletion to address the following issues:
1. **Handling Complex Morphological Structures**: Real-world objects, such as plants, often have diverse morphological structures and discontinuous surfaces, which make point cloud completion more challenging.
2. **Limitations of Existing Methods**: Existing point cloud completion methods assume that objects in the input dataset have regular shapes and continuous surfaces, an assumption that does not always hold in practical applications.
3. **Improving Completion Performance**: Existing methods perform poorly when dealing with complex morphological structures, necessitating a new approach to improve completion accuracy.

To tackle these issues, the paper introduces morphological segmentation, which segments the point set through deep hierarchical feature learning, breaking down complex morphological structures into regular shapes and continuous surfaces. Additionally, the paper leverages HDBSCAN (a density-based clustering algorithm) to effectively identify point cloud instances belonging to the same feature type and uses a multiscale generative network to achieve feature point-based missing point cloud repair. To compensate for the average distance difference between the patch center and its nearest neighbors, the paper introduces a simple yet effective uniform loss.

Experimental results show that SegCompletion outperforms existing methods in handling 3D shape point cloud completion tasks with different morphologies.