| 摘 要: 针对目前点云采样方法的细节缺失问题,提出了一种基于定位细化的多层3D点云上采样方法,以稀疏点云作为输入,实现高质量点云上采样。该方法主要由粗糙点云生成和定位细化两个部分组成。在粗糙点云生成阶段,利用基于平面折叠的自编码器架构将输入的稀疏三维点云进行粗糙重构;在定位细化阶段,利用细化单元与偏移回归机制对粗糙点云进行位置修正,引入局部细化与全局细化模块以学习更精细的几何结构。与已有同类方法比较,所提方法不但在形状采样任务上取得了更好的成效,而且在形状补全任务上也展现出优秀的能力。 |
| 关键词: 点云上采样 平面折叠编码器 定位细化 偏移回归 |
|
中图分类号: TP391.41
文献标识码: A
|
|
| Research on 3D Point Cloud Upsampling Method Based on Position Refinement |
|
HE Long, WU Zhihe, LIU Wujie, KUANG Zhenzhong
|
(School of Computer, Hangzhou Dianzi University, Hangzhou 310018, China)
lhe@hdu.edu.cn; wuzhihe01@hdu.edu.cn; liuwujie@hdu.edu.cn; zzkuang@hdu.edu.cn
|
| Abstract: To address the issue of detail loss in current point cloud upsampling methods, a mult-i layer 3D point cloud upsampling method based on position refinement is proposed. This method takes a sparse point cloud as input and achieves high-quality point cloud upsampling. It primarily consists of two components: coarse point cloud generation and position refinement. In the coarse point cloud generation stage, a plane-folding-based autoencoder architecture is employed to roughly reconstruct the input sparse 3D point cloud. In the position refinement stage,
refinement units and an offset regression mechanism are used to correct the positions of the coarse point cloud. Local and global refinement modules are introduced to learn finer geometric structures. Compared with existing methods, the proposed approach not only achieves better performance in shape upsampling tasks but also demonstrates excellent capability in shape completion tasks. |
| Keywords: point cloud upsampling plane folding encoder position refinement offset regression |
