Abstract
3D line skeletons are simplistic representations of a shape’s topology which are used for a wide variety of geometry-processing tasks, including shape recognition, retrieval, and reconstruction. Numerous methods have
been proposed to generate a skeleton from a given 3D shape. While mesh-based methods can exploit existing knowledge about the shape’s topology and orientation, point-based techniques often resort to precomputed per-
point normals to ensure robustness. In contrast, previously proposed techniques for unprocessed point clouds either exhibit inferior robustness or require expensive operations, which in turn increases computation time. In this
paper, we present a new and highly efficient skeletonization approach for raw point cloud data, which produces overall competitive results compared to previous work, while exhibiting much lower computation times. Our algo-
rithm performs robustly in the face of noisy and fragmented inputs, as they are usually obtained from real-world 3D scans. We achieve this by first transferring the input point cloud into a Gaussian mixture model (GMM), obtaining
a more compact representation of the surface. Our method then iteratively projects a small subset of the points into local L1-medians, yielding a rough outline of the shape’s skeleton. Finally, we present a new branch detection
technique to obtain a coherent line skeleton from those projected points. We demonstrate the capabilities of our proposed method by extracting the line skeletons of a diverse selection of input shapes and evaluating their visual
appearance as well as the efficiency compared to alternative state-of-the-art methods
been proposed to generate a skeleton from a given 3D shape. While mesh-based methods can exploit existing knowledge about the shape’s topology and orientation, point-based techniques often resort to precomputed per-
point normals to ensure robustness. In contrast, previously proposed techniques for unprocessed point clouds either exhibit inferior robustness or require expensive operations, which in turn increases computation time. In this
paper, we present a new and highly efficient skeletonization approach for raw point cloud data, which produces overall competitive results compared to previous work, while exhibiting much lower computation times. Our algo-
rithm performs robustly in the face of noisy and fragmented inputs, as they are usually obtained from real-world 3D scans. We achieve this by first transferring the input point cloud into a Gaussian mixture model (GMM), obtaining
a more compact representation of the surface. Our method then iteratively projects a small subset of the points into local L1-medians, yielding a rough outline of the shape’s skeleton. Finally, we present a new branch detection
technique to obtain a coherent line skeleton from those projected points. We demonstrate the capabilities of our proposed method by extracting the line skeletons of a diverse selection of input shapes and evaluating their visual
appearance as well as the efficiency compared to alternative state-of-the-art methods
Original language | English |
---|---|
Pages (from-to) | 38-47 |
Number of pages | 10 |
Journal | Journal of WSCG |
Volume | 30 |
Issue number | 1-2 |
DOIs | |
Publication status | Published - 2022 |
Event | 30th International Conference on Computer Graphics, Visualization and Computer Vision: WSCG 2022 - Pilsen, Czech Republic Duration: 17 May 2022 → 20 Sept 2022 https://www.wscg.cz/ |
Keywords
- curve skeleton
- Gaussian mixture
- geometric computation
- point cloud
ASJC Scopus subject areas
- Software
- Computational Mathematics
- Computer Graphics and Computer-Aided Design