Semantic Segmentation for Point Cloud based on Distance Weighted and Adaptive Augmentation

Since a scene represented with point clouds usually have richer information compared with 2D image and is capable of achieving an intuitive reflection of the real world, the analysis of original point cloud data has attracted intensive research interests in recent years. Due to the disordered nature of 3D data, perceptually understanding of large-scale 3D data is very challenging, especially for the dense prediction task such as the semantic segmentation. As limited by the computation resources, most existing approaches are only able to deal with small patches that are extracted from a large-scale 3D data, which is still a problem. In this work, we concentrate on the essential visual task, semantic segmentation, for large-scale point cloud data collected in reality. On the one hand, to reduce the ambiguity in nearby points, we augment their local context by features and spatial distance mutually weighted. On the other hand, to reduce the loss of information during the aggregation of neighborhood features, we introduce an attention mechanism to adaptively enhance local features. Further, we perform semantic segmentation on benchmarks S3DIS and our method has achieved outstanding performance.