SCA-PointNet++: Fusion Spatial-Channel Attention for Point Cloud Extraction Network of Buildings in Outdoor Scenes

The development of acquisition and generation techniques for 3D point clouds of large-scale outdoor scenes provides a prerequisite for subsequent digital twin analysis of the 3D world. However, due to the irregularity and complexity of 3D point cloud data, extracting buildings directly from large-scale point clouds is still full of challenges. In this paper, based on the introduction of Convolutional Attention Mechanism (CBAM) for PointNet++ network, and feature coding techniques, SCA-PointNet++ fusion channel-space attention point cloud extraction network for buildings in outdoor scenes is proposed. Enhanced local spatial feature saliency through local spatial location coding, channel-spatial attention mechanism to improve the learning ability of salient structures, and disparity pooling under the attention mechanism to improve the feature delivery efficiency and better point cloud segmentation are evaluated on the SensatUrban dataset. A comparative study with five established deep learning point cloud classification models confirms that our proposed SCA-PointN et++ achieves good performance in the task of classifying cloud buildings at outdoor site locations.