Point Cloud Sampling Preserving Local Geometry for Surface Reconstruction
Kohei Matsuzaki (KDDI Research, Inc.),* Keisuke Nonaka (KDDI Research, Inc.)
The 34th British Machine Vision Conference
Abstract
Surface reconstruction from point clouds is a fundamental task in computer vision and graphics. Recent methods learn neural fields as a surface representation from point clouds. However, these methods are difficult to scale to large scenes due to the limited size of the point clouds they can handle. In this paper, we propose a point cloud sampling method to improve scalability for training a surface reconstruction network. We train the surface reconstruction network with sampled point clouds obtained from a sampling network. In the sampling network, we introduce a seed point that serves as the origin to sample point clouds from partial regions. It encourages the surface reconstruction network to learn both the global structure and local geometry on a part of the scene. We also introduce a split-and-merge approach to avoid increasing the memory footprint by suppressing the input size to the sampling network. Experimental results on ScanNet dataset show that the proposed method significantly improves surface reconstruction performance compared with state-of-the-art methods.Video
Citation
@inproceedings{Matsuzaki_2023_BMVC,
author = {Kohei Matsuzaki and Keisuke Nonaka},
title = {Point Cloud Sampling Preserving Local Geometry for Surface Reconstruction},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0306.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by