**5. Conclusions and Future Work**

In this study, a DGCNN was used to learn indoor 3D point cloud features. Five items, i.e., columns, beams, walls, floors, and ceilings, were considered the research objects. Training and learning were conducted using two different sources of point cloud data: S3DIS and photogrammetry. Area\_1 data in the S3DIS dataset and those of corridor 2F in the civil engineering building were used as test samples for segmentation.

The endpoint feature extraction program developed in this study was applied to process the segmented results. Other details, such as endpoint coordinates, quantity, and length of the research object, are derived. Finally, feature information is imported into an automatic modeling program for parametric element modeling.

This study uses a DGCNN to learn the features of indoor point clouds and segment the point clouds of columns, beams, walls, floors, and ceilings automatically. The overall accuracies using the S3DIS indoor dataset and civil engineering building information were 86.9% and 94.2%, respectively.

An endpoint feature extraction method that overcomes the errors caused by irregular line segments is proposed in this paper. In addition, for columns and beams with low semantic segmentation accuracy, a range processing method is devised to reduce semantic segmentation errors.

The method can be employed to calculate the number of components, boundary length and size, and relative information from the extracted endpoint. In comparing the inspected size of corridor 2F in the civil engineering building with the measurement yielded by automatic modeling, the RMSE is found to be approximately ±0.03 m. Because the point clouds are constrained by control points, the model is similar to the building.

The results of this study demonstrate that indoor 3D point clouds produced by closerange images can be segmented using a trained 3D deep learning network. The automatic feature point extraction method proposed in this study is employed to derive the feature point information of components. Using this information, the point cloud can be imported into an automatic modeling system to generate BIM parametric components and create indoor drawings.

There are a lot of objects in the room, and this study only sets out to study and discuss five categories of structural objects. In the future, we intend to increase the number of samples, increase the types of objects, reduce the noise, and explore ways to improve accuracy.

**Author Contributions:** Conceptualization, funding acquisition, methodology, writing—original draft preparation, C.-S.H. and X.-J.R.; data curation, software (data experiments), writing—review and editing, visualization, C.-S.H. and X.-J.R.; formal analysis, investigation (physical experiments), validation, X.-J.R. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Data Availability Statement:** This research did not use publicly archived datasets.

**Conflicts of Interest:** The authors declare no conflict of interest.
