融合多尺度上下文的点云语义分割方法OA

Current research on semantic segmentation based on point cloud largely ignores the global context information and lacks the ability to extract point feature.To address the problem,a point cloud semantic segmentation network that combines feature representation augmentation and global context aggregation was proposed.Firstly,it introduces more geometric and feature contexts from explicit 3D space and implicit feature space to the point cloud,and improves the dependency between point descriptors and channel descriptors through bilinear regularization,retaining sufficient global information.Secondly,it combines with dilated convolution to expand the receptive field of the point cloud,enhancing the feature extraction capability.Finally,it fuses multi-scale features of different encoding layers to realize the complementarity of geometric information and semantic feature.Comparative experiments conducted on S3DIS and Semantic3D indicate that mean intersection over union of our proposed method reaches 71.9％and 78.1％respectively,which are 1.9％and 0.7％higher than that of the baseline network RandLA-Net.