Continuous conditional random field convolution for point cloud segmentation
作者：
Highlights：
• A continuous CRF graph convolution (CRFConv) is proposed to model the upsampling process of point cloud features.
• A point cloud segmentation network based on the CRFConv is built to enhance the location ability of the network.
• The classical discrete CRF is also reformulated as an additional graph convolution to refine the labels.
• A dual CRF network is implemented to model the data affinity in both feature space and label space simultaneously.
• The experiments on various challenging benchmarks demonstrate the effectiveness and robustness of the proposed method.
摘要
•A continuous CRF graph convolution (CRFConv) is proposed to model the upsampling process of point cloud features.•A point cloud segmentation network based on the CRFConv is built to enhance the location ability of the network.•The classical discrete CRF is also reformulated as an additional graph convolution to refine the labels.•A dual CRF network is implemented to model the data affinity in both feature space and label space simultaneously.•The experiments on various challenging benchmarks demonstrate the effectiveness and robustness of the proposed method.
论文关键词：Point cloud segmentation,Conditional random fields,Message passing,Graph convolution,Mean-field approximation
论文评审过程：Received 12 August 2020, Revised 6 August 2021, Accepted 27 September 2021, Available online 28 September 2021, Version of Record 10 October 2021.
论文官网地址：https://doi.org/10.1016/j.patcog.2021.108357
