A Hierarchical Neural Network for Point Cloud Segmentation and Geometric Primitive Fitting

Automated generation of geometric models from point cloud data holds significant importance in the field of computer vision and has expansive applications, such as shape modeling and object recognition. However, prevalent methods exhibit accuracy issues. In this study, we introduce a novel hierarchi...

Full description

Saved in:
Bibliographic Details
Published in:Entropy (Basel, Switzerland) Switzerland), 2024-08, Vol.26 (9), p.717
Main Authors: Wan, Honghui, Zhao, Feiyu
Format: Article
Language:English
Subjects:
Analysis
Computer vision
Editing
Geospatial data
Image segmentation
Machine vision
Neural networks
Noise levels
Noise prediction
Object recognition
Parameter estimation
Parameter identification
Parameter robustness
point cloud
primitive fitting
segmentation
Semantics
Subdivisions
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Automated generation of geometric models from point cloud data holds significant importance in the field of computer vision and has expansive applications, such as shape modeling and object recognition. However, prevalent methods exhibit accuracy issues. In this study, we introduce a novel hierarchical neural network that utilizes recursive PointConv operations on nested subdivisions of point sets. This network effectively extracts features, segments point clouds, and accurately identifies and computes parameters of regular geometric primitives with notable resilience to noise. On fine-grained primitive detection, our approach outperforms Supervised Primitive Fitting Network (SPFN) by 18.5% and Cascaded Primitive Fitting Network (CPFN) by 11.2%. Additionally, our approach consistently maintains low absolute errors in parameter prediction across varying noise levels in the point cloud data. Our experiments validate the robustness of our proposed method and establish its superiority relative to other methodologies in the extant literature.
ISSN:1099-4300
1099-4300
DOI:10.3390/e26090717