2. Geometric distribution of point cloud

The point cloud data of walls, floors, and ceilings has a flat and wide distribution. In contrast, the point cloud geometry of columns and beams has a 90◦ corner.

#### *4.3. Semantic Segmentation and Modeling*

After completing segmentation, the result of each test area was imported into the feature extraction operation. Moreover, parametric modeling was performed using automatic modeling rules.

### 4.3.1. S3DIS Modeling

In the test sample of S3DIS, three sections of Area\_1 were selected for the analysis of segmentation results: Conference\_Room2, Office\_2, and Office\_6. Feature extraction and parametric modeling of the three areas were performed sequentially. The modeling results are shown in Figure 5 to verify the feasibility of the automatic modeling rule design. To facilitate the visualization of internal modeling, the ceiling is removed, as shown in Figure 5.

**Figure 5.** Automatic modeling results of (**a**) Conference\_Room2, (**b**) Office\_2, and (**c**) Office\_6 without ceiling.

#### 4.3.2. 2F Corridor of Civil Engineering Building

After the segmentation of the point cloud in corridor 2F, feature extraction and automatic modeling were performed sequentially. The modeling results are shown in Figure 6. The ceiling is also removed to visualize the interior.

**Figure 6.** Automated modeling results of corridor 2F (without ceiling) in civil engineering building.

#### *4.4. Evaluation of 3D Model*

As presented in this section, the modeling results of corridor 2F in the civil engineering building of our school were selected for testing. This is because corridor 2F is more convenient to measure on site than the other areas. There were two columns, four walls, two beams, one floor, and a ceiling in the area. This study analyzes the top view and sectional view, as shown in Figure 7.

To verify whether the modeling result is consistent with the actual length of the selected area, a total station was used for measuring the points, as shown in Figure 8. The comparison results are listed in Table 7. The root mean square error (RMSE) of the line length given by BIM compared with the actual length is ±0.03 m.

**Figure 7.** (**a**) Top view and (**b**) sectional view of corridor 2F.

**Figure 8.** Line position map for comparing 3D model and in situ field. The letters A–Q indicate the location of the points, and the length of the line segment connecting the two points is measured and compared in this study.

**Table 7.** Three-dimensional model and actual line length difference (unit: m).


The feature extraction process developed in this study derives the features of endpoints from the point cloud segmentation results. Then, the parametric elements of columns, beams, walls, floors, and ceilings were automatically modeled based on the attribute information. The experimental results indicate that Area1\_ConferenceRoom2, Area1\_Office 2, Area1\_Office 6 of S3DIS, and corridor 2F in the civil engineering building can be used to create the 3D model data of indoor components automatically. Each component has attribute information, such as material, length, volume, and quantity. The overall process not only reduces the time cost of manual model construction but also serves as follow-up application management. It is a rapid BIM method for reconstructing existing indoor spaces.

In comparing the actual length (obtained by inspection) of corridor 2F with the indoor measurement yielded by automatic modeling, the RMSE is found to be ±0.03 m; hence, the accuracy is acceptable.
