The document describes an approach for automatically segmenting 3D point cloud scans of indoor environments into rooms and detecting openings between rooms. It uses the initial assignment of points to scans to generate an initial room segmentation. It then formulates refining the segmentation as a probabilistic clustering problem based on the "average visibility" of points belonging to different rooms. Detected planar surfaces are used to find intersections between scans indicating openings. The approach is demonstrated on several real-world point cloud datasets, showing its ability to generate structural building descriptions that could enable applications like facilities management. Limitations in handling non-convex rooms and potential solutions using "reachability" are also discussed.

1. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Automatic Generation of Structural Building
Descriptions from 3D Point Cloud Scans
GRAPP 2014 – Paper ID 54
Sebastian Ochmann
ochmann@cs.uni-bonn.de
University of Bonn, Germany
January 6th, 2014
Future Work

2. Introduction
Related Work
Room Segmentation
Door Detection
Introduction
Results & Conclusion
Future Work

3. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Strong trend in architecture towards Building Information Modeling
(BIM) for planning, facility management, and retrofitting purposes.
• In addition to geometry also includes meta data and entity
relations.
• BIM models not readily available for older buildings, still
desirable, e.g. for renovation planning.
Image from http://www.digital210king.org/

4. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Strong trend in architecture towards Building Information Modeling
(BIM) for planning, facility management, and retrofitting purposes.
• In addition to geometry also includes meta data and entity
relations.
• BIM models not readily available for older buildings, still
desirable, e.g. for renovation planning.
Image from http://www.digital210king.org/

5. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Strong trend in architecture towards Building Information Modeling
(BIM) for planning, facility management, and retrofitting purposes.
• In addition to geometry also includes meta data and entity
relations.
• BIM models not readily available for older buildings, still
desirable, e.g. for renovation planning.
Image from http://www.digital210king.org/

6. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Increasing availability and usage of laser scans as a starting
point/guide for BIM generation.
• Point clouds lack structure, making BIM generation a highly
manual, time-consuming process.
• Automatic methods for structural and semantic analysis of
point clouds are essential.
Image from http://www.digital210king.org/

7. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Increasing availability and usage of laser scans as a starting
point/guide for BIM generation.
• Point clouds lack structure, making BIM generation a highly
manual, time-consuming process.
• Automatic methods for structural and semantic analysis of
point clouds are essential.
Image from http://www.digital210king.org/

8. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Increasing availability and usage of laser scans as a starting
point/guide for BIM generation.
• Point clouds lack structure, making BIM generation a highly
manual, time-consuming process.
• Automatic methods for structural and semantic analysis of
point clouds are essential.
Image from http://www.digital210king.org/

9. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
This paper presents an approach for room segmentation and
opening detection from indoor point clouds.
• Facilitates navigation within and handling of point clouds,
enables highlighting/hiding of individual rooms.
• Automatic placement of doors, approximation of room areas.
• Enables retrieval of room constellations (graph queries).
Partial scan of Kronborg castle (Denmark); room segmentation and detected connections between rooms are shown.

10. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
This paper presents an approach for room segmentation and
opening detection from indoor point clouds.
• Facilitates navigation within and handling of point clouds,
enables highlighting/hiding of individual rooms.
• Automatic placement of doors, approximation of room areas.
• Enables retrieval of room constellations (graph queries).
Partial scan of Kronborg castle (Denmark); room segmentation and detected connections between rooms are shown.

11. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
This paper presents an approach for room segmentation and
opening detection from indoor point clouds.
• Facilitates navigation within and handling of point clouds,
enables highlighting/hiding of individual rooms.
• Automatic placement of doors, approximation of room areas.
• Enables retrieval of room constellations (graph queries).
Partial scan of Kronborg castle (Denmark); room segmentation and detected connections between rooms are shown.

12. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
This paper presents an approach for room segmentation and
opening detection from indoor point clouds.
• Facilitates navigation within and handling of point clouds,
enables highlighting/hiding of individual rooms.
• Automatic placement of doors, approximation of room areas.
• Enables retrieval of room constellations (graph queries).
Partial scan of Kronborg castle (Denmark); room segmentation and detected connections between rooms are shown.

13. Introduction
Related Work
Room Segmentation
Door Detection
Related Work
Results & Conclusion
Future Work

14. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Turner and Zakhor: Floor Plan Generation and Room Labeling of
Indoor Environments from Laser Range Data (GRAPP 2014 –
yesterday)
Image by Turner and Zakhor
• Generation of triangulated floor plan from 2D or 3D point cloud.
• Room labeling formulated as graph-cut problem.
• Generation of 2.5D, watertight models with room segmentation.
N.B.: This paper is not included in the related work of our paper as it was published after submission deadline.

15. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Mura et al.: Robust Reconstruction of Interior Building Structures
with Multiple Rooms under Clutter and Occlusions
(CAD/Graphics, November 2013)
Image by Mura et al.
• Generation of 2D cell complex from wall candidates.
• Diffusion embedding of 2D cell complex for clustering rooms.
N.B.: This paper is not included in the related work of our paper as it was published after submission deadline.

16. Introduction
Related Work
Room Segmentation
Door Detection
Room Segmentation
Results & Conclusion
Future Work

17. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
First goal: Segmentation of point cloud into rooms.
Future Work

18. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Starting point: Multiple separate, registered point cloud scans,
including scanner positions.
Idea: Use point-to-scanner assignments as initial, coarse “guess”
for room segmentation.

19. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Idea: Resolve incorrect labelings by determining which room labels
are most visible from a certain point.
E.g., a “red” point inside of the “green” room is likely to be part
of the green room because it “sees” mostly green points.

20. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Note on initial “point-to-scan” labeling:
Scans belonging to the same room need to be merged.
Currently done manually; automatic merging suggestions may be
given (see paper).

21. Introduction
Related Work
Room Segmentation
Door Detection
Note on visibility measure between points
Results & Conclusion
Future Work

22. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Detect planar structures1 with (smoothed) occupancy bitmaps.
1
Schnabel et al.: Efficient RANSAC for Point-Cloud Shape Detection (2007).
Future Work

23. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Estimate visibility (value in [0, 1]) between two points by testing
for intersections with the planes.

24. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Let vj (xk ) be an “average” visibility from point k to all points
currently assigned to room j (see paper for details).
Room j
Point k
Future Work

25. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Formulation of relabeling as probabilistic clustering problem.
Future Work

26. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Formulation of relabeling as probabilistic clustering problem.
Future Work

27. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Formulation of relabeling as probabilistic clustering problem.
Future Work

28. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Formulation of relabeling as probabilistic clustering problem.
Room prior (governed by "room size")
Future Work

29. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Formulation of relabeling as probabilistic clustering problem.
Room prior (governed by "room size")
Class-conditional probability of k'th point
(governed by "average visibility" of room)

30. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Formulation of relabeling as probabilistic clustering problem.
Room prior (governed by "room size")
Class-conditional probability of k'th point
(governed by "average visibility" of room)
Class-cond. prob.
Room prior

31. Introduction
Related Work
Room Segmentation
Door Detection
Example for iterative relabeling procedure.
Iteration 0/7 (initial situation)
Results & Conclusion
Future Work

32. Introduction
Related Work
Room Segmentation
Door Detection
Example for iterative relabeling procedure.
Iteration 1/7
Results & Conclusion
Future Work

33. Introduction
Related Work
Room Segmentation
Door Detection
Example for iterative relabeling procedure.
Iteration 2/7
Results & Conclusion
Future Work

34. Introduction
Related Work
Room Segmentation
Door Detection
Example for iterative relabeling procedure.
Iteration 3/7
Results & Conclusion
Future Work

35. Introduction
Related Work
Room Segmentation
Door Detection
Example for iterative relabeling procedure.
Iteration 4/7
Results & Conclusion
Future Work

36. Introduction
Related Work
Room Segmentation
Door Detection
Example for iterative relabeling procedure.
Iteration 5/7
Results & Conclusion
Future Work

37. Introduction
Related Work
Room Segmentation
Door Detection
Example for iterative relabeling procedure.
Iteration 6/7
Results & Conclusion
Future Work

38. Introduction
Related Work
Room Segmentation
Door Detection
Example for iterative relabeling procedure.
Iteration 7/7
Results & Conclusion
Future Work

39. Introduction
Related Work
Room Segmentation
Door Detection
Door Detection
Results & Conclusion
Future Work

40. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Second goal: Find openings (e.g. doors) between adjacent rooms;
construct room connectivity graph, e.g. for enabling retrieval.

41. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Observation: Openings cause overlaps between scans.
Future Work

42. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
The relabeling process has just resolved these overlaps.
Future Work

43. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Extract scanner-to-point rays corresponding to relabeled points.
Future Work

44. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Determine intersection points of rays with detected planes.
Future Work

45. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Extract pairs of intersection points to approximate openings.
Estimated door
position and size
Future Work

46. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Results & Conclusion
Future Work

47. Introduction
Related Work
Room Segmentation
Door Detection
5 scans of Kronborg castle, Denmark.
Results & Conclusion
Future Work

48. Introduction
Related Work
Room Segmentation
Door Detection
5 scans of Kronborg castle, Denmark.
Results & Conclusion
Future Work

49. Introduction
Related Work
Room Segmentation
Door Detection
6 scans of Kronborg castle, Denmark.
Results & Conclusion
Future Work

50. Introduction
Related Work
Room Segmentation
Door Detection
6 scans of Kronborg castle, Denmark.
Results & Conclusion
Future Work

51. Introduction
Related Work
Room Segmentation
Door Detection
14 scans of a building in Denmark.
Results & Conclusion
Future Work

52. Introduction
Related Work
Room Segmentation
Door Detection
14 scans of a building in Denmark.
Results & Conclusion
Future Work

53. Introduction
Related Work
Room Segmentation
Door Detection
14 scans of a building in Denmark.
Results & Conclusion
Future Work

54. Introduction
Related Work
Room Segmentation
Door Detection
14 scans of a building in Denmark.
Results & Conclusion
Future Work

55. Introduction
Related Work
Room Segmentation
Door Detection
14 scans of Oslo bispeg˚rd.
a
Results & Conclusion
Future Work

56. Introduction
Related Work
Room Segmentation
Door Detection
14 scans of Oslo bispeg˚rd.
a
Results & Conclusion
Future Work

57. Introduction
Related Work
Room Segmentation
Door Detection
14 scans of Oslo bispeg˚rd.
a
Results & Conclusion
Future Work

58. Introduction
Related Work
Room Segmentation
Door Detection
14 scans of Oslo bispeg˚rd.
a
Results & Conclusion
Future Work

59. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Limitations: Highly non-convex rooms cause problems.
Part of Risløkka trafikkstasjon, Oslo.
Future Work

60. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Problem:
• Assumption that (almost) all points of a room are visible from
any point within that room is violated.
Possible solution (also see next slides):
• Use (possibly indirect) “reachability” instead of visibility.
• Take into account not only direct line-of-sight but also
indirect connections, allowing to “see around corners”.

61. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Problem:
• Assumption that (almost) all points of a room are visible from
any point within that room is violated.
Possible solution (also see next slides):
• Use (possibly indirect) “reachability” instead of visibility.
• Take into account not only direct line-of-sight but also
indirect connections, allowing to “see around corners”.

62. Introduction
Related Work
Room Segmentation
Door Detection
Future Work
Results & Conclusion
Future Work

63. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Future Work
Improved room segmentation, also works on non-convex datasets.

64. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Note: Opening detection is not restricted to doors.
Future Work

65. Introduction
Related Work
Room Segmentation
Door Detection
Results & Conclusion
Note: Opening detection is not restricted to doors.
Future Work

66. ACKNOWLEDGEMENTS
We would like to thank Henrik Leander Evers for the scans
of Kronborg Castle, Denmark, and the Faculty of Architecture and Landscape Sciences of Leibniz University Hannover for providing the 3D building models that were used
for generating the synthetic data. This work was partially
funded by the European Communitys Seventh Framework
Programme (FP7/2007-2013) under grant agreement no.
600908 (DURAARK) 2013-2016