Automatic extraction and reconstruction of a 3D wireframe of an indoor scene from semantic point clouds

Figures & data

Figure 1. Flowchart of the proposed method.

Figure 2. Floor boundary primitive extraction. (a) Rough boundary of ${\alpha }_1$. (b) Detailed boundary of ${\alpha }_2$. (c) Illustration of combined polygons. (d) Result of boundary extraction.

Figure 3. Wall primitives extraction with the M-RSC method.

Figure 4. Illustration of the boundary primitives alignment.

Figure 5. Classification of feature points.

Figure 6. Illustration of linear feature optimization. (a) Perpendicular condition. (b) Collinear condition.

Figure 7. Pipeline of the iterative semantic constrained optimization method.

Figure 8. Comparison between the initial 2D wireframe and optimized 2D wireframe.

Figure 9. Pipeline of door and window wireframe extraction.

Figure 10. Illustration of height information extraction. (a) Schematic diagram of the cylinder method. (b) Example of height extraction by the cylinder method.

Figure 11. Height information optimization. (a) Abnormal height. (b) Height undulation. (c) Corrected abnormal height. (d) Smooth height edges.

Figure 12. Reference height extraction by a height histogram. (a) Histogram of the height distribution of points. (b) Corresponding plane detected by the histogram.

Figure 13. Height information extraction and reconstruction. (a) Height information in the point cloud. (b) Reconstruction structure from height information.

Figure 14. Three experimental scenes.

Table 1. Data description of testing scenes.

Indoor scene	Scene 1	Scene 2	Scene 3
Source	S3DIS Dataset	ISPRS Dataset	ISPRS Dataset
Category	Long Corridor	Multiroom	Granger Museum
Characteristic	Long and narrow, many doors and windows	High clutter with a height change	High clutter and complex
Point number	4481021	8688375	23517095
Clutter	Low	Medium	High
Area ($m^2$)	88.882	238.996	2020.340

Table 2. Parameters for floor boundary extraction.

Indoor scene	Scene 1	Scene 2	Scene 3
${\mathit{\alpha }}_1$(m)	15	15	15
${\mathit{\alpha }}_2$(m)	0.1	0.1	0.1

Table 3. Parameters for wall boundary extraction.

Indoor scene	Scene 1	Scene 2	Scene 3
R	0.03	0.05	0.03
$N_{stop}$	5	15	10

Figure 15. Three indicators of accuracy evaluation.

Figure 16. Initial 2D wireframe of experimental scenes. (a) Scene 1. (b) Scene 2. (c) Scene 3.

Figure 17. Optimized 2D wireframe of experimental scenes. (a) Scene 1. (b) Scene 2. (c) Scene 3.

Figure 18. 3D wireframes of Scene 1. (a) 3D wireframe with semantics. (b) Details of doors and windows.

Figure 19. 3D wireframe of Scene 2.

Figure 20. 3D wireframe of Scene 3.

Figure 21. Distribution of the accuracy evaluation. (a) Planar accuracy. (b) Curve accuracy. (c) Height accuracy.

Table 4. Average of evaluation indicators.

Indicator1	$\mathrm{\Delta }d$(m)	Improvement
Our wireframe	0.044	20%
Initial wireframe	0.055
Indicator2	$\mathrm{\Delta }\alpha {(}^{\circ })$	Improvement
Our wireframe	2.197	33.3%
Initial wireframe	3.298
Indicator3	$\mathrm{\Delta }h$(m)	Improvement
Our wireframe	0.008	87.7%
BIM	0.065

Figure 22. 2D Method comparison on Scene 3. (a) BIM. (b) our method. (c) Wu et al.’s method.

Figure 23. 3D Method comparison on Scene 2. (a) Our method. (b) Oesau et al.’s method. (c) Wang et al.’s method. (d) Han et al.’s method.

Table 5. The position error of points of Scene 2.

Method	Ours	Wang et al.’s	Han et al.’s
Position error(m)	0.050	0.211	0.142

Table 6. The position error of points of Scene 3.

Method	Ours	Wu et al.’s
Position error(m)	0.069	0.087

Figure 24. Semantic segmentation results optimization. (a) semantic point cloud with error segmentation. (b) semantic segmentation point cloud after processing.

Figure 25. Illustration of ${\alpha }_1$ selection.

Figure 26. Illustration of non-horizontal structure. (a) the sloping roof that bulges in the middle of the building. (b) curved roofs.