.png){kind=icon}
.png){kind=icon}
Return
PU-GACNet: Graph Attention Convolution Network for Point Cloud Upsampling
.png){kind=icon}
PRE
AI
DOI：10.1016/j.imavis.2021.104371.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
ABSTR A C T Real-scanned point clouds are often sparse and non-uniform. To conquer the problem, researchers propose point cloud upsampling techniques, whose efficiency and effectiveness heavily rely on their feature extractors and fea-ture expanders used therein. Therefore, in this paper, to capture the global and local structured features of point clouds, we first design a Graph Attention Convolution (GAC) module as a feature extractor by assigning different attentional weights to combine spatial positions and feature attributes dynamically. Furthermore, we propose an Edge-aware NodeShuffie (ENS) module as a feature expander to upsampling point features smoothly, in an effort to better preserve local geometric details and emphasize local edges. Finally, we combine GAC module with ENS module into a novel point cloud upsampling pipeline, named as PU-GACNet. Extensive experiments as well as theoretical analysis demonstrate this pipeline significantly outperforms previous methods in network perfor-mance for 3D point cloud upsampling, to obtain more efficient inference with much fewer parameters.(c) 2022 Elsevier B.V. All rights reserved.
Keywords:
Point cloud upsampling
Graph attention convolution
Feature extraction
Edge-aware nodeshuffie
Feature expansion
Journal
IF:
4.2
Papers:
4.0K
Citations:
6.7K