Abstract: In order to improve the accuracy of point cloud semantic segmentation in indoor scenes, an all fusion  network for semantic segmentation of 3D point clound is proposed. The network consists of a feature encoding module, a progressive feature decoding module, a multi-scale feature decoding module, a feature fusion module, and a semantic segmentation header. The feature encoding module uses inverse density weighted convolution as the feature encoder to perform hierarchical feature encoding on point cloud, so as to extract multi-scale features of the point cloud. Then, the progressive feature decoder is used to decode high-level semantic features layer by layer to obtain the point cloud progressive decoding feature. In the same pair, the multi-scale feature decoder performs feature decoding on the extracted multi-scale features to obtain multi-scale decoding features of the point cloud. Finally, the progressive decoding feature is fused with the multi-scale decoding feature, then semantic segmentation header is introduced to realize the point cloud semantic segmentation. The all fusion network robustly enhances the feature extraction ability of the network, and the experimental results also verify the effectiveness of the method.

Key words: all fusion network, feature fusion, semantic segmentation, 3D point cloud, deep learning