Point Cloud Overlapping Region Co-Segmentation Network

Kexue Fu, Xiaoyuan Luo, Manning Wang
NeurIPS 2020 Workshop on Pre-registration in Machine Learning, PMLR 148:1-13, 2021.

Abstract

3D point clouds are being increasingly used in the field of computer vision and many applications involve the processing of partially overlapping point clouds. However, little attention has been paid to the property of partial overlap. In this paper, we propose the concept of co-segmentation of the overlapping region of two 3D point clouds and develop a deep neural network to solve this problem. The proposed network utilizes co-attention mechanism to aggregate information from the paring point clouds so as to find the overlapping region. The co-segmentation of overlapping region can be regarded as a preprocessing step in practical 3D point cloud processing pipelines so that downstream tasks can be better accomplished. We build a dataset of partially overlapping 3D point clouds from ModelNet40 and ShapeNet, which are two widely used 3D point cloud datasets, and the overlapping region can be obtained automatically without manual labelling. We also utilize the real 3D point cloud datasets, 3DMatch and ScanNet, in which the overlapping region can be obtained from the relative pose between point clouds provided in the datasets. We evaluate the performance of the proposed method on co-segmentation of overlapping region on these datasets and its effectiveness in improving one downstream task, 3D point cloud registration, which is very sensitive to partial overlapping

Cite this Paper

BibTeX
@InProceedings{pmlr-v148-fu21a, title = {Point Cloud Overlapping Region Co-Segmentation Network}, author = {Fu, Kexue and Luo, Xiaoyuan and Wang, Manning}, booktitle = {NeurIPS 2020 Workshop on Pre-registration in Machine Learning}, pages = {1--13}, year = {2021}, editor = {Bertinetto, Luca and Henriques, João F. and Albanie, Samuel and Paganini, Michela and Varol, Gül}, volume = {148}, series = {Proceedings of Machine Learning Research}, month = {11 Dec}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v148/fu21a/fu21a.pdf}, url = {https://proceedings.mlr.press/v148/fu21a.html}, abstract = {3D point clouds are being increasingly used in the field of computer vision and many applications involve the processing of partially overlapping point clouds. However, little attention has been paid to the property of partial overlap. In this paper, we propose the concept of co-segmentation of the overlapping region of two 3D point clouds and develop a deep neural network to solve this problem. The proposed network utilizes co-attention mechanism to aggregate information from the paring point clouds so as to find the overlapping region. The co-segmentation of overlapping region can be regarded as a preprocessing step in practical 3D point cloud processing pipelines so that downstream tasks can be better accomplished. We build a dataset of partially overlapping 3D point clouds from ModelNet40 and ShapeNet, which are two widely used 3D point cloud datasets, and the overlapping region can be obtained automatically without manual labelling. We also utilize the real 3D point cloud datasets, 3DMatch and ScanNet, in which the overlapping region can be obtained from the relative pose between point clouds provided in the datasets. We evaluate the performance of the proposed method on co-segmentation of overlapping region on these datasets and its effectiveness in improving one downstream task, 3D point cloud registration, which is very sensitive to partial overlapping} }
Endnote
%0 Conference Paper %T Point Cloud Overlapping Region Co-Segmentation Network %A Kexue Fu %A Xiaoyuan Luo %A Manning Wang %B NeurIPS 2020 Workshop on Pre-registration in Machine Learning %C Proceedings of Machine Learning Research %D 2021 %E Luca Bertinetto %E João F. Henriques %E Samuel Albanie %E Michela Paganini %E Gül Varol %F pmlr-v148-fu21a %I PMLR %P 1--13 %U https://proceedings.mlr.press/v148/fu21a.html %V 148 %X 3D point clouds are being increasingly used in the field of computer vision and many applications involve the processing of partially overlapping point clouds. However, little attention has been paid to the property of partial overlap. In this paper, we propose the concept of co-segmentation of the overlapping region of two 3D point clouds and develop a deep neural network to solve this problem. The proposed network utilizes co-attention mechanism to aggregate information from the paring point clouds so as to find the overlapping region. The co-segmentation of overlapping region can be regarded as a preprocessing step in practical 3D point cloud processing pipelines so that downstream tasks can be better accomplished. We build a dataset of partially overlapping 3D point clouds from ModelNet40 and ShapeNet, which are two widely used 3D point cloud datasets, and the overlapping region can be obtained automatically without manual labelling. We also utilize the real 3D point cloud datasets, 3DMatch and ScanNet, in which the overlapping region can be obtained from the relative pose between point clouds provided in the datasets. We evaluate the performance of the proposed method on co-segmentation of overlapping region on these datasets and its effectiveness in improving one downstream task, 3D point cloud registration, which is very sensitive to partial overlapping
APA
Fu, K., Luo, X. & Wang, M.. (2021). Point Cloud Overlapping Region Co-Segmentation Network. NeurIPS 2020 Workshop on Pre-registration in Machine Learning, in Proceedings of Machine Learning Research 148:1-13 Available from https://proceedings.mlr.press/v148/fu21a.html.

Related Material