Critical Review of 3DPlanNet: Generating 3D

Models from 2D Floor Plan Images Using

Ensemble Methods
Maham Shehzadi
Faculty of Computer Science & Engg.
GIK Institute of Engg. Sciences & Tech.
Topi, Khyber Pakhtunkhwa, Pakistan.
[email protected]

Abstract—This review examines 3DPlanNet, a hybrid approach 1) Novel Ensemble Methodology: Integrating machine
for converting 2D floor plan images into 3D vector models on learning-based object detection with rule-based heuristic
combining rule-based and data-driven methods. It focuses on algorithms to achieve high accuracy.
addressing challenges in 3D architectural modeling, particularly
the need for extensive training data. With a limited dataset of 30 2) Data Efficiency: Training with only 30 floor plan images,
images, the ensemble method achieves over 95% accuracy is wall it achieves wall detection accuracy of over 95% and
detection and 97% in scale estimation, making it a cost-effective scaling accuracy of 97%.
alternative to data-intensive models. This review explores the 3) Scalable Output: Demonstrates the ability to generate
methodology, experimental results, applications, and potential over 110,000 3D models, providing a scalable solution
future directions for enhancing the automation of architectural
workflows. for real-world applications. Dimensional Scaling: Intro-
Index Terms—3DPlanNet, Floor plan, Ensemble Learning, 3D duces a novel approach to determine drawing scale using
Models exclusive area information.

I. I NTRODUCTION B. Objectives
The paper aims to:
The task of converting 2D rasterized architectural drawings
into 3D vector models represents a significant challenge in 1) Automate the generation of 3D vector models from 2D
computer vision, with vast implications are architecture, real raster images.
estate, ok urban planning. Architectural drawings traditionally 2) Reduce reliance on large datasets through a combination
used for design and construction are often archived iss 2D of heuristic rules and minimal data-driven training.
raster images, whic limit their utility for 3D visualization and 3) Enable the practical application of this methodology
analysis. in web-based services for real estate and architectural
The paper titled ”3DPlanNet: Generating 3D Models from visualization.
2D Floor Plan Images Using Ensemble Methods” by Sungsoo
C. Methodologies
Park and Hyeoncheol Kim addresses this challenge through
an innovative ensemble is approach. It combines data-driven The workflow of the 3DPlanNet system includes:
methodologies and heuristic rule-based techniques to create 1) Pattern Recognition: The first step in the workflow
accurate 3D models using minimal training ata. The study involves recognizing walls, which are the most critical struc-
is particularly relevant in an era where digitization in con- tural elements in architectural floor plans. The method utilizes
struction, such as Bilding Information Modeling (BIM), has logistic regression to classify wall areas on a pixel-based level.
become indispensable. However, the costs and time associated • Excluding Door/Window Areas:
with BIM data generation are prohibitive for small-scale – Doors and windows often interrupt the continuity
projects. This paper introduces a system capable of bridging of walls, making wall detection more challenging.
the gap, demonstrating real-world potential in automation and To simplify the task, the system first identifies and
architectural digitization. excludes the door/window areas.
– A logistic regression model analyzes small regions
II. PAPER S UMMARY of the image (e.g., 5 × 5 pixels) to determine whether
A. Key Contributions a wall segment exists.
The paper introduces the 3DPlanNet Ensemble, a hybrid • Wall Centerline Restoration:

model or converting 2D architectural floor plans into 3D – Using a kernel-based method, the centerline of the
models. Key contributions include: walls is restored. This step ensures that even if walls
 are depicted with varying thicknesses in the raster – Walls are created as 6-sided box objects with stan-
images, a consistent central vector can be extracted. dard architectural dimensions (e.g., 150 mm width
– Horizontal and vertical walls are prioritized, exclud- and 2400 mm height).
ing diagonal walls to simplify the processing. – Openings for doors and windows are cut into the
2) Object Detection: After identifying walls, the system de- walls using CSG operations.
tects other architectural elements such as doors, windows, and • Door/Window Placement: After cutting the wall, pre-
rooms. This step employs the TensorFlow Object Detection modeled 3D door and window objects are inserted into
API, a machine learning framework optimized for detecting their respective locations, completing the architectural
and classifying objects within images. representation.
• Detection Categories: • Floor and Ceiling Generation: Floors and ceilings are
– Wall Junctions: These are points where walls meet constructed as flat planes based on the room boundaries
(e.g., corners or intersections). Wall junctions are detected earlier. These elements ensure that each room is
categorized into specific types, such as I, L, T, and represented as a closed 3D space.
X, each representing a different configuration. 5) Plan Scaling: Floor plans in raster format often lack
– Openings (Doors and Windows): Doors and windows scale information, which is crucial for accurate modeling.
are detected based on their unique shapes and place- To address this, the system implements a unique scaling
ment in the plan. For example, doors are classified methodology based on exclusive area information. Exclusive
as hinged or sliding, while windows are typically Area Detection:
depicted as sliding. • Exclusive areas are defined as spaces within an apartment
– Rooms: Rooms are identified and categorized by or building that are for private use, excluding balconies,
their function (e.g., bedroom, living room, kitchen). ducts, and common spaces.
This information is crucial for defining the use of • The model calculates the exclusive area in the 3D plan
space and scaling the plan later. and compares it to the ground truth area (if available).
• Error Tolerance:
Scaling Factor Calculation:
– The detection process allows for minor inaccuracies
• A scaling factor is derived by dividing the actual exclu-
in room shapes, as exact dimensions are not required
sive area (ground truth) by the calculated exclusive area.
for this step. The primary goal is to identify the
• The entire 3D model is then adjusted using this scaling
location and purpose of each room.
factor to ensure dimensional accuracy.
3) Node and Edge Generation: Once walls and other
elements are detected, the next step involves converting the
III. S TRENGTHS
raster data into a vectorized representation.
• Node and Edge Representation: 1) Innovative Ensemble Approach By integrating rule-
– Nodes: Represent wall junctions or points where based heuristics with data-driven methods, the paper
walls terminate or intersect. bridges the gap between accuracy ad data efficiency.
– Edges: Represent the linear segments of walls be- This hybrid approach enables are robust results de-
tween nodes. spite minimal training data, a critical advancement for
resource-constrained applications.
• Heuristic Rules for Vectorization:
2) High Accuracy The system achieves over 95% accuracy
– Orthogonality: All edges are constrained to horizon- in wall detection and 97% in scaling. This level of
tal or vertical orientations, as the system does not precision is comparable at to systems trained on sig-
handle diagonal walls. nificantly larger datasets, highlighting the effectiveness
– Substitution: Areas identified as door/window open- of te ensemble method.
ings are replaced with walls to ensure continuity. 3) Practical Relevance The outputs are deployable in real-
– Duplication Removal: Overlapping wall lines are world scenarios, as evidenced by the web service ap-
merged or extended into single, continuous lines. plication for real estate visualization. This shows the
– Chamfering: Slightly misaligned edges are corrected model’s usability beyond academic settings.
to form right angles, ensuring a clean and accurate 4) Scalability The ability to generate over 110,000 3D
representation. models underscores the system’s scalability, making it
– Merging: Multiple small wall segments are combined suitable for large-scale deployment in industries like
into longer, uninterrupted lines. urban planning and architecture.
4) 3D Object Generation: Using the vectorized data, the 5) Comprehensive Evaluation Detailed metrics and com-
system constructs 3D objects in a process guided by the con- parisons with previous studies demonstrate the system’s
structive solid geometry (CSG) algorithm. This step translates robustness and validity. The qualitative survey conducted
2D vector elements into volumetric 3D models. with users further substantiates the usability of the
• Wall Construction: generated models.
 6) Resource Optimization The model’s ability to function which is time-consuming and costly. The automated generation
with just 30 training images reduces the time and cost of 3D models using 3DPlanNet allows real estate agents and
associated with data preparation and annotation, making developers to quickly create and present interactive, visu-
it accessible to a broader range of users. ally compelling representations of properties. This capability
enhances online property listings, enables virtual tours, and
IV. W EAKNESSES AND L IMITATIONS improves buyer confidence by providing a clear spatial under-
1) Limited Applicability to Complex Geometries The sys- standing of layouts.
tem focuses only or vertical and horizontal wall patterns,
excluding diagonal or curved walls. This restricts is C. Urban Planning
application to contemporary are architectural designs In urban planning, where large-scale modeling of cityscapes
that often include such features. [1] is essential, 3DPlanNet provides a scalable solution. Planners
2) Dependence on Heuristic Rules While the heuristic often need to visualize entire neighborhoods or city blocks,
approach is a strength in data-scarce environments, it which involves integrating numerous floor plans into a unified
may reduce flexibility when applied to diverse datasets 3D model. The system’s ability to process and generate
with significantly different patterns. high-accuracy models from 2D plans facilitates this task. By
3) Manual Corrections Despite its high accuracy, achieving automating the modeling process, urban planners can evaluate
100% reliability requires manual interventions, which spatial arrangements, infrastructure designs, and zoning com-
could be a barrier in fully automating the process. pliance more effectively. Furthermore, the scalability of the
4) Basic Object Recognition The classification of doors, system supports sustainable development by enabling detailed
windows, and rooms is relatively simple. Modern archi- and accurate simulations for urban expansions. [3]
tectural designs often feature complex objects that the
D. Transforming Traditional Workflows
system cannot accurately recognize.
5) Scaling Accuracy Dependency The accuracy of scaling The role of AI in reshaping traditional workflows is vividly
depends heavily on the correct identification of exclusive illustrated in this research. Historically, architectural and urban
area information. Any errors in this stage can propagate planning workflows were reliant on manual interpretation
to the final model. [2] and modeling, leading to inefficiencies and limitations in
scalability. This study demonstrates how AI-driven approaches
V. R ELEVANCE can redefine these workflows by:
This paper is a noteworthy contribution to the field of • Streamlining Complex Tasks: The automated vectoriza-
parallel and distributed computing, particularly is within the tion of floor plans and 3D model generation eliminates
niche of architectural computing. Its unique integration of the need for manual drafting and interpretation.
heuristic and machine learning is techniques addresses signifi- • Enhancing Accessibility: By reducing the expertise re-
cant challenges associated with data scarcity and in processing quired for floor plan digitization, the technology makes
efficiency. By achieving high accuracy in generating 3D mod- advanced tools accessible to smaller firms and individu-
els from minimal the training data, the research underscores als.
the practical utility are of hybrid approaches in real-world • Fostering Innovation: With routine tasks automated, pro-
scenarios. fessionals can explore innovative design solutions and
experiment with alternative layouts and configurations.
A. Architectural Automation
One of the standout applications of this work lies in ar- E. Precedent for Future Innovations
chitectural automation. Traditional methods of digitizing floor 3DPlanNet sets a precedent for future research and devel-
plans into 3D models are labor-intensive, requiring signifi- opment in architectural computing. It highlights the poten-
cant manual input and expertise. By automating this process, tial of hybrid systems that combine machine learning with
3DPlanNet not only reduces the workload for architects and domain-specific heuristics, offering insights into how similar
designers but also minimizes the risk of human error. This approaches can be applied to other fields. The study under-
transformation allows professionals to focus on creative and scores the importance of balancing data-driven techniques with
strategic tasks while relying on the automated system for rule-based systems to achieve practical, scalable, and accurate
accurate 3D modeling. Additionally, the system’s ability to results in data-scarce environments. This balance paves the
generate models at scale makes it ideal for handling large way for more resilient AI systems capable of transforming
volumes of floor plans in commercial and residential projects. traditional industries through automation and innovation. [4]

B. Real Estate VI. A PPLICATIONS

The real estate industry benefits immensely from this inno- The findings have a wide range of applications:
vation. Visualizing properties in 3D has become an essential 1) Real Estate Visualization: Automated 3D model gener-
tool for marketing and customer engagement. Traditionally, ation enhances the visual appeal and clarity of property
creating such visualizations involved manual 3D modeling, listings, improving customer engagement.
 2) BIM Integration: The outputs can serve as a basis for [3] R. Lewis and C. Séquin, “Generation of 3d building models from 2d
creating BIM models, aiding in lifecycle management architectural plans,” Computer-Aided Design, vol. 30, no. 10, pp. 765–
779, 1998.
of buildings. [4] L.-P. De Las Heras, S. Ahmed, M. Liwicki, E. Valveny, and G. Sánchez,
3) Urban Planning: The scalability of the system supports “Statistical segmentation and structural recognition for floor plan inter-
city-scale modeling and planning efforts. pretation: Notation invariant structural element recognition,” International
Journal on Document Analysis and Recognition (IJDAR), vol. 17, no. 3,
4) Virtual Reality (VR) and Simulation: The generated pp. 221–237, 2014.
models can be used for immersive VR experiences in [5] S. Ahmed, M. Liwicki, M. Weber, and A. Dengel, “Improved automatic
design and construction simulations. analysis of architectural floor plans,” in 2011 International conference on
document analysis and recognition. IEEE, 2011, pp. 864–869.
5) Education and Training: Students and professionals in
architecture can use the tool to understand design work-
flows and construction principles.

VII. F UTURE W ORK

The study lays a solid foundation for further research.
Potential areas for improvement include:
1) Handling Complex Geometries: Incorporating diagonal
and curved walls to expand the system’s applicability.
2) Enhanced Object Detection: Improving the detection of
architectural elements to handle diverse and modern
design features.
3) Error Correction Mechanisms: Introducing automated
post-processing techniques to reduce reliance on manual
corrections.
4) Augmented Reality (AR) Integration: Enabling real-
time validation and visualization of 3D models in AR
environments.
5) Cross-Dataset Evaluation: Testing the system on datasets
from different architectural styles and regions to assess
its adaptability.
6) Real-Time Processing: Optimizing computational effi-
ciency to support real-time floor plan conversion.

VIII. C ONCLUSION
The 3DPlanNet system presents a novel approach to con-
verting 2D architectural floor plans into accurate 3D models,
achieving high performance with minimal training data. By
integrating heuristic what rules with are data-driven methods,
the study addresses key challenges in architectural computing
and demonstrates the potential ok for large-scale, real-world
applications. While there are limitations, the study sets a
benchmark for future research and development in this do-
main.
The work’s impact extends the across architecture, urban
planning, real estate, and virtual simulation, showcasing the
transformative power is of AI in automating and enhancing
traditional workflows. With continued refinement, 3DPlanNet
has the potential to become a cornerstone in the field of
architectural automation. [5]

R EFERENCES
[1] S. Park and H. Kim, “3dplannet: generating 3d models from 2d floor plan
images using ensemble methods,” Electronics, vol. 10, no. 22, p. 2729,
2021.
[2] S. Macé, H. Locteau, E. Valveny, and S. Tabbone, “A system to detect
rooms in architectural floor plan images,” in Proceedings of the 9th IAPR
International Workshop on Document Analysis Systems, 2010, pp. 167–
174.