Analysis and Evaluation of PointNet for Indoor Office Point Cloud Semantic Segmentation

Abstract

Indoor modeling is one of the primary sources of information in building management due to the increased use of BIM in the AEC industry. The indoor model can be acquired with several survey instruments, but TLS is the most popular resulting point cloud that can be processed into a 3D model. However, the process commonly still uses inefficient manual methods. Point cloud data have irregular, unordered, unstructured characteristics, making them more challenging to process. The deep learning algorithm can be a solution to solve the problem. PointNet is the first deep learning algorithm that directly accepts point cloud data as input. This study aims to analyze and evaluate the office indoor point cloud segmentation using PointNet. The office indoor point cloud data was acquired using TLS and then pre-processed for deep learning input. Transfer learning strategy is used as a weight initialization technique. The pre-trained model was trained with the S3DIS dataset and then fine-tuned to segment nine indoor classes in this study. The result shows PointNet achieves 85% overall accuracy and 66% average class IoU score to predict indoor classes using this study’s point cloud data. Geometry control shows that the predicted point cloud has an RMSE score of 1.8 cm, meaning the geometries of the segmented point cloud are accurate. Using the transfer learning method has increased the performance of the deep learning model. Further research is needed to evaluate the model thoroughly using more training and evaluation data and different transfer learning strategies.