.png){kind=icon}
.png){kind=icon}
Return
MTCloud: Multi-type convolutional linkage network for point cloud instance segmentation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.126432.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Semantic and instance segmentation of point clouds are essential tasks of 3D scene understanding, which are widely used in various fields, including autonomous driving, robotics, augmented reality, and high-definition urban mapping. Semantic segmentation aims to assign semantic labels to each point. Instance segmentation involves segmenting different objects with the same semantic class. Therefore, semantic segmentation can be considered a foundational task for instance segmentation. Based on this motivation, we design a point cloud instance segmentation network named MTCloud, based on a multi-type convolutional linkage algorithm. First, we propose the Contextual Sparse Convolution Module to efficiently extract the three-dimensional geometric structure features of the point cloud, which performs neighborhood feature aggregation on the sparse tensor. This approach reduces memory consumption by avoiding the processing of empty voxels. We also introduce a Sparse Voxel Transformer Module that utilizes sparse convolution to capture both global and local features of the point clouds. Moreover, the proposed method mitigates the limitation of transformers in processing only a small number of points per pass, improving the capability to capture comprehensive point cloud features. Additionally, we combine spatially sparse convolution with submanifold sparse convolution in the Spatial Submanifold Dilated Convolution Module. This module uses different dilation values to expand the network's receptive field while preserving the point cloud's sparsity. We conducted extensive experiments on indoor point cloud datasets ScanNet v2 and S3DIS, as well as the 3D aerial photogrammetry point cloud dataset STPLS3D, demonstrating its compatibility with existing instance segmentation frameworks. The results show that MTCloud improves instance segmentation performance by enhancing semantic feature extraction, achieving significant gains compared to the original instance segmentation models. MTCloud serves as a practical and effective approach to bridging semantic and instance segmentation tasks.
Keywords:
Multi-type convolutional linkage
Point cloud
Semantic segmentation
Instance segmentation
Deep learning
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W
Organization
No organization information available