.png){kind=icon}
.png){kind=icon}
Return
LCL_FDA: Local context learning and full-level decoder aggregation network for large-scale point cloud semantic segmentation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2024.129321.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Point cloud semantic segmentation is a key techno-logy for autonomous driving, intelligent power inspection, and smart city planning, etc. However, many existing methods still face limitations in capturing local contextual information and learning long-range dependencies in point cloud representations. To adder-ss these problems, we propose a novel local context learning and full-level aggregation network for point cloud semantic segmentation in large-scale scenes. Firstly, we construct a local geometry feature extraction module using both coordinate and angular information. This module aims to acquire global information within local regions, thereby enhancing locally discriminative features. Secondly, we combine single-point channel attention pooling and rotation axis cross attention pooling to effectively enhance the capability of network to exploit local details. Finally, we devised a full-level decoder aggregation module, which expands the receptive field of each decoder to enhance the learning capability of the network for long-range dependencies. We evaluate our method on three large-scale point cloud datasets, including our self-constructed DW_3D dataset and the publicly available datasets Toronto3D and S3DIS. The experiment results validate the effectiveness of our method. The mIoU of the proposed method on DW_3D, Toronto-3D and S3DIS dataset can reach 72.9 %, 82.2 %, 72.3 %(6-fold). Code is avaiable https://github.com/yzq-stu/LCL_FDA.
Keywords:
Attention mechanism
Large-scale point cloud
Point cloud dataset in power scene
Semantic segmentation
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W