Subdivision of Adjacent Areas for 3D Point Cloud Semantic Segmentation

In 3D point cloud semantic segmentation, much of the previous research has focused on aggregating the fine-grained geometric structures of local regions, overlooking the long-term features. However, global long-term contextual dependencies play a role as important as local features aggregation. This paper proposes a Subdivision of Adjacent Areas (SAA) module, which efficiently mines more informative features to enrich global long-term contextual dependencies. SAA is constructed by the CoVariance-Enhanced Channel Attention (CECA) and the PseudoNL Spatial Attention (PSA). The former learns the interdependence among channels via second-order statistics for each feature channel, while the latter efficiently captures the positional correlation among points in the entire space via a pseudo feature map. The proposed SAA, a plug-and-play, end-to-end trainable module, can be integrated into existing segmentation networks. Extensive experiments on S3DIS and ScanNet datasets demonstrate that networks integrated with our SAA improve mIoU performance. It verifies that SAA is beneficial for 3D point cloud segmentation networks in achieving excellent performance.