Semantic Segmentation of 3D Building Facade Point Clouds Using Dynamic Graph CNN

Abstract

The 3D point cloud representation of building holds significant importance in developing a comprehensive smart city model. However, due to the substantial volume of point cloud data and its inherently unstructured nature, accomplishing semantic segmentation of the point cloud poses a formidable challenge. Recent years have witnessed substantial advancements in employing deep learning techniques for point cloud segmentation. Both PointNet and PointNet++ have demonstrated their prowess in managing point cloud data. In the context of this study, we harness the potential of DGCNN for segmenting building point clouds. DGCNN plays a pivotal role in partitioning the input point cloud, thereby influencing the sensor domain's network range. Consequently, this research delves into the role of parameter k in K-NN (K-Nearest Neighbors) and its implications. The findings illustrate that augmenting the 'k' value contributes to enhancing the overall precision of DGCNN. Nonetheless, complete precision in segmenting every module cannot be assured. The research objectives in the study is building façade, and the targets are windows, walls, balconies and doors. In this study, the key parameters are the values of 'k' and the 'block size.' Experiments were conducted by varying the values of these two parameters to obtain different segmentation results of building facade point clouds. Notably, setting k to 20, block size to 1m achieves an overall accuracy rate of 89.75%, a mean accuracy rate of 85.03%, an IoU of 76.24%, while elevating k to 30 and block size to 1m results in a heightened accuracy rate of 95.74%, a mean accuracy rate of 89.37%, an IoU of 81.52%.