Developing a Robust Training Dataset for AI-Driven Architectural Spatial Layout Generation

Abstract

Recent advancements in AI research, particularly in spatial layout generation, highlight its capacity to enhance human creativity by swiftly providing architects with numerous alternatives during the pre-design phase. The complexity of architectural design data, characterized by multifaceted elements and varying representations, presents significant challenges in creating uniform and robust datasets. This study addresses this challenge by developing a robust training dataset specifically tailored for AI-driven spatial layout generation in architecture. An algorithm capable of extracting spatial relationship diagrams from raster-based floor plan images and converting them into vector-based data was introduced. Through extensive web crawling, a dataset comprising 10,000 data rows, categorized into 21 classes and three spatial relationship categories, was collected. When tested with the You-Only-Look-Once (YOLO) model, the detection rate was 99%, the mean average precision was 85%, and the MIoU was 74.2%. The development of this robust training dataset holds significant potential to advance knowledge-based artificial intelligence design automation studies, paving the way for further innovation in architectural design.

1. Introduction

Searle, a renowned American philosopher and academic, distinguished between “weak” and “strong” artificial intelligence (AI) that simulates human cognitive capacities using computers [1]. The advent of AI research in architectural design signifies more than its mere utilization as an instrument; it suggests a progression toward “strong AI” that enhances intrinsic design creativity. Recently, many studies on spatial layout generation [2,3,4,5] and recommendations [6] have been conducted using AI technology. A commonality between these studies is that they can increase human creativity by quickly providing architects with many alternatives during the predesign phase. Such AI-based research in the field of architecture and the construction of datasets is necessary and must precede [7].

However, architectural design data are replete with intricate elements and the representations in drawings often vary. Consequently, the uniform formulation of datasets is challenging and circumscribed. Considering that superior datasets directly influence the outcomes of deep learning, it is imperative to prioritize the construction of extensive high-quality training datasets. Therefore, this paper proposes a methodology and algorithm for constructing high-quality datasets geared toward AI research in spatial layout generation. The main objective of this study is to develop a robust training dataset specifically tailored for AI-driven spatial layout generation in architecture. By achieving this, the study aims to support the advancement of knowledge-based AI design automation.

By consistently extracting extensive information from floor plan images (such as space names, areas, and adjacencies) as vector data, knowledge-based AI learning beyond intelligent learning is possible. This approach holds promise for evolving into a knowledge-based AI design automation technique.

2. Related Research

2.1. AI-Based Dataset Construction for Architectural Design

One reason for the slow application and development of AI in architectural design is the difficulty of constructing standardized datasets with diverse data. The Korea Institute of Architecture and Urban Research [8] aggregates and offers information related to design projects, including drawings, images, and summaries. However, the volume of this data falls short of the AI learning requirements. The floor plan is expressed using heuristic symbols, including lines, guides, text, and icons, which can be interpreted as meaningful information. However, a universally recognized procedure for the transition from raster to vector data is still lacking, complicating efforts to achieve uniformity [9,10].

For effective AI learning, the datasets must be constructed in a format amenable to computer processing. Given these datasets, the AI can learn spatial rules to generate or recommend an optimized layout. Lu [11] proposed an algorithm for generating graphs using CubiCase5K, floor plan datasets. Based on the BIM model and Dynamo, Kim [12] interpreted the spatial relationship information in a drawing as a graph and constructed a spatial relationship database using an algorithm. In the early stages of design education, Kavakoglu [13] used sketch datasets expressed by students to train the styleGAN2-ADA model and generate facade images. Pizarro [14] investigated the datasets (such as name, public access, annotation, and number of plans) that were mainly used for floor plan analysis and recognition.

Notably, design-centric datasets are remarkably small compared to those in other fields. This is limited in annotation because it is difficult to obtain raster-based raw data and there are many heuristics in drawings. Specifically, the levels and details of the raw data are different; therefore, when annotating, the classes, boundaries, and standards are inevitably different. Consequently, high-quality large-scale training datasets should be constructed in accordance with specific research objectives and orientations. This study distinguishes itself by not only addressing these common challenges but also by developing a novel algorithm that consistently extracts extensive information from floor plan images and converts it into vector-based data, specifically tailored for spatial layout generation in architecture.

2.2. AI-Based Spatial Layout Generation Research

Hopgood [15] defines AI as the intelligence of a machine or computer that enables it to imitate or mimic human capabilities. Regarding this perspective, AI-based tools are divided into “intelligent systems” or “knowledge-based systems” and “computational intelligence and hybrid systems”. Intelligent systems are highly relevant in design automation [16].

This study delves into recent deep-learning spatial layout generation studies in the field of computer vision. Chaillou [2] trained a model to generate a footprint, proceeded with a generative adversarial network (GAN)-generated master plan for program repartitioning, and proposed a furniture layout. After the floor plans, doors, windows, and vertical circulations were placed, a layout was created using GAN learning. Nauata [3] studied a house GAN layout that used an architectural constraint as a graph (i.e., the number and types of rooms with their spatial adjacency) and produced a set of axis-aligned bounding boxes of rooms. Wu [4] determined the location of a room as a point given a boundary and predicted the location of a wall using an encoder–decoder model. By repeatedly learning to create different rooms, this study constructed a vector dataset for more than 80,000 residential floor plans. Carta [17] utilized the CubiGraph5k [11] dataset. Utilizing machine-parsing structured floor plan data and Graph Convolutional Networks (GCN), an approximate spatial layout is generated when a user draws rooms (kitchen, living room, etc.) as nodes like a graph. Seo [18] learned drawing elements and spatial usage using the DeeplabV3+ model and then automatically generated spatial zoning based on AttnGAN. Hu [5] learned to generate floor plans from layout graphs using the RPLAN’s dataset [4]. Given an input building boundary, a floor plan associated with the layout graphs is generated from a database.

A unifying aspect of prior studies is a rule-based randomly generated floor plan. However, these studies did not consider design theory or knowledge. Although AI-based intelligent systems boast rapid result generation, their results often lack precision and are unreliable. For better quality and more reliable plans, it should be developed as an AI-based knowledge system with high reliability. This necessitates the construction of training datasets consistently rooted in vector data.

The training datasets should include complex vector data, such as type, size, scale, location, and adjacency of space, rather than simply drawing elements, such as walls, doors, windows, and furniture. This is because the more detailed the vector data about space that are constructed, the better the AI can learn these spatial relationships through neural networks. Moreover, the current study differentiates itself by introducing a novel algorithm designed to convert raster-based floor plan images to vector-based data using deep learning, providing a strategy for formulating expansive training datasets tailored for high-precision AI applications in architectural design. Such insights will be instrumental in generating or recommending spaces in the future.

3. Methodology

3.1. Procedure

The procedure adopted in this study is illustrated in Figure 1. Initially, raster-based floor plan images served as the raw data sources. Subsequently, the classes and adjacency conditions for the space were defined. In the third phase, training data were created using an annotation tool and data training and testing were conducted using the You-Only-Look-Once (YOLO) recognition model. Finally, an algorithm for automatically constructing vector-based data from raster-based floor plan images was proposed and tested. The dataset, named ‘AIBIM_House Dataset’ [19], comprises graph-extracted spatial relationship diagrams converted into vector-based data. Metrics such as precision, detection rate, mean Average Precision (mAP), and Intersection over Union (IoU) were used to measure the algorithm’s performance.

3.2. Raw Data Collection

Raw data were collected by crawling raster-based floor plan images from real estate websites. Web crawling can be used to collect and archive specific data [20]. The total number of projects on ‘Houseplans’ is 14,970, classified by style, size, favorites, and region [21]. Each project was assigned a distinct identity (ID) and included a raster-based elevation plan, floor plan, and text-based description. Web crawling, executed using Python, facilitated the raw data collection process, as illustrated in Figure 2. By inputting the project ID, specific content could be retrieved (selected from Cases 1, 2, and 3). From 6900 projects, 10,000 raster-based floor plan images (in formats such as .jpg and .png) and their corresponding descriptions (in .xlsx format) were collected. These 10,000 images underwent spatial relationship diagram extraction using YOLO-based recognition. Finally, text-based descriptions and spatial relationship diagrams were algorithmically converged to extract the final vector-based spatial relationship data.

3.3. Spatial Relationship Definition

Before delving into spatial relationships, categorizing housing spaces is essential. According to architectural planning theories, housing spaces are typically classified into indoor and outdoor areas. Indoor spaces encompass public, private, household, sanitary, and circulation spaces. By comparing floor plan images from Korea and the United States, representative categories of housing spaces were identified. Appendix A provides a detailed hierarchy of these spaces. In this study, we defined 14 indoor and 2 outdoor space classes. The indoor space categories include living rooms, dining rooms, bedrooms, dressing rooms, closets, kitchens, utility rooms, pantry rooms, laundry rooms, bathrooms, entrances, stairs, and hallways. The distinction between stair rooms and entrances was made for annotation purposes. The two outdoor space categories are porches and garages. As depicted in

Table 1. Definition of classes of housing space.

Division		Class Name	Class Number	Color
Indoor space	Public space	Living room	1	RGB(254,0,2)
		Dining room	20	RGB(254,0,2)
	Private space	Bedroom	2	RGB(255,255,0)
		Dressing room	5	RGB(255,255,0)
		Closet	11	RGB(255,255,0)
	Household space	Kitchen	3	RGB(0,0,254)
		Utility room	6	RGB(0,0,254)
		Pantry room	9	RGB(0,0,254)
		Laundry room	10	RGB(0,0,254)
	Sanitary space	Bathroom	4	RGB(0,176,80)
	Aisle space	Entrance	0	RGB(128,15,181)
		Stair room	7	RGB(128,15,181)
		Stair entrance	17	RGB(128,15,181)
		Hallway	8	RGB(128,15,181)
Outdoor space		Porch	18	RGB(76, 76, 76)
		Garage	19	RGB(76, 76, 76)

, housing space classes were defined for image recognition.

In the pre-design phase, spatial relationships can be represented in graphs, particularly in space programming and zoning. These graphs are denoted by nodes that symbolize the spaces and the edges that interconnect them. Pizarro [14] suggested considering four categories when analyzing floor plans: graphic separation, object recognition, vectorization, and structural modeling. Zeng [22] used a multitask network model to recognize floor plan elements such as walls, rooms, doors, and windows. Song [23] used image-based learning to classify floor plan elements and represented them in a vector format. Their methodology emphasizes the detection of indoor elements (such as walls and doors), symbols, and spatial elements (such as rooms and corridors) for image learning.

Based on prior studies, it is necessary to detect space and basic elements (walls, doors, and windows) to analyze spatial relationships in the floor plan. As shown in

Table 2. Definition of Adjacency in Spatial Relationships.

Floor plan
Definition of Adjacency in Spatial Relationships
Directly connected spaces	Case 1		No walls (Completely open space)
	Case 2		Walls but no doors (Partially open space)
Indirectly connected space	Case 3		Walls and door or window
Disconnected space	-		Only walls

, this study defines three cases for determining adjacency in spatial relationships. Case 1 is a directly connected space, that is, a completely open space without walls. Case 2 was a directly connected space but was a partially open space with walls and no doors. Case 3 is an indirectly connected space with walls and a door or window. A space with walls alone is disconnected. The criteria for discerning adjacency within spatial relationships are based on elements, such as doors (single, double, or sliding), windows, and openings. Thus, this study defined 16 spatial classes alongside five central classes and established three distinct cases to elucidate the adjacencies in spatial relationships.

3.4. YOLO-Based Spatial Relationship Diagram Extraction Method

The convolutional neural network (CNN) and YOLO [24,25] are pivotal object detection models. Although object detection requires the ability to identify and categorize multiple entities, CNN models encounter challenges when recognizing numerous objects. As floorplan images contain a variety of diminutive and intricate elements of varying sizes and positions, the YOLO model was adopted in this study. This decision stems from YOLO’s proficiency in rapidly detecting multiple classes within raster-based floor plan images. Unlike Fast R-CNN, which requires two stages, specifically region proposal and classification, YOLO models are single-stage detectors. This means that YOLO can predict bounding boxes and class probabilities directly from full images in one evaluation, making it faster and more suitable for real-time applications. YOLOv3, in particular, predicts the objectness score for each bounding box using logistic regression. YOLOv3’s architecture divides the input image into a grid and predicts bounding boxes and confidence scores for objects within each grid cell. This offers significant advantages in terms of both speed and accuracy, especially when detecting smaller objects compared to other models like Faster R-CNN. In terms of operational efficiency, YOLOv3 outperforms both Faster R-CNN and Retinanet and its ability to detect smaller objects has been enhanced using scales [25].

Table 3. Extraction process of the spatial relationship diagram.

Step 1. Image input		Step 2. Class detection using YOLO model		Step 3. Center point calculated from bounding box
	→		→
Step 4. Connection by definition of adjacency			Step 5. Spatial relationship diagram extraction
		→

presents the extraction process of the spatial relationship diagram from YOLOv3-based floor plan images. When an image was input, the space and central classes were detected using the YOLO model. A center point is detected and each point (referring to a node) is connected based on previously defined adjacency conditions (i.e., a completely open space, a partially open space, or an indirectly connected space). This is discussed in detail in Section 3.3.

Table 4. Connection line according to the adjacency definition.

Division	Case 1			Case 2			Case 3
Definition	Completely open space			Partially open space			Indirectly connected space
Example
connection line
		Thick line			Thin line			Dotted line

shows how the lines were connected according to the adjacency definition.

Table 5. Encapsulation methods for extracting spatial relationship diagrams using the YOLO model.

Step 1. Detected Bounding Box		Step 2. Offset for Overlapping Regions		Step 3. Convert to Square and Determine Diagram Size		Step 4. Calculated Center Point		Step 5. Connect and Extract Circular Diagram
	→		→		→		→

lists the successful encapsulation methods used in the extraction process. It is divided into a step for detecting classes using the YOLO model and a step for drawing a diagram, such as a graph, through calculation. The dimensions of the graph images were determined after data calculation. If the outer coordinates of the detected space class are represented by X and Y, their cumulative lengths are incorporated and averaged to obtain a square. As the space classes are discerned through bounding boxes, overlaps may occur. In the event of such overlaps, the principle of pushing each other was applied. The points of minimum X and Y and maximum X and Y were found to connect these four points with a line to create a boundary. After determining the central coordinates within the square, the connections between them were drawn based on their relative adjacencies. The final diagrammatic representation was obtained by transforming the square into a circular format. Consequently, a spatial relationship diagram is extracted from the original raster-based floor plan images.

4. Algorithm and Experiment

4.1. YOLO-Based Training Dataset Construction Algorithm

4.1.1. Algorithm for Spatial Relationship Diagram Extraction

The algorithms and contents for extracting spatial relationship diagrams are depicted in Figure 3 and

Table 6. Description of Figure 3.

Procedure	Meaning of Input	Meaning of Output
Select floor plan → Search same project	Select one floor plan image	Search the same project as the selected floor plan image
Detection floor plan → Search class	Detect floor plan image	Search space and medium class
Connectalorithm.cs → Connection information	Run ‘Connectalorithm.cs’ to determine the adjacency of spatial relationships	Know connected information
Drawimages.cs → Show diagram image	Run‘Drawimages.cs’	Show spatial relationship diagram image

. Each project is uniquely identified by an ID. The term ‘Search Same Project’ means to search for the same project from the selected floor plan image (for instance, Project ID-3 corresponds to a building on the 3rd floor).

The intricacies and procedures within ‘Connectalorithm.cs’ are exhibited in Figure 4. Within housing plans, living rooms predominantly occupy the most extensive area. Consequently, directly adjacent spaces like hallways, kitchens, and foyers are the primary connections. Rooms with a solitary door are connected accordingly. In instances where a room contains multiple doors, all neighboring rooms are interconnected. As space is detected as a bounding box, the closet is mostly included in the bedroom. The porch space is connected and the hallways may overlap due to bounding boxes, so they are merged. The ‘Connectorithm.cs’ function yields ‘Connection Information’ as its output. The contents described in Chapter 3.4 are in ‘Drawimages.cs’. The conclusive diagrammatic image is presented to the user after calibrating parameters like pixels, resolution, and dimensions.

4.1.2. Algorithm for Spatial Relationship Data Extraction

The algorithm and contents for extraction of spatial relationship data are shown in Figure 5. The conclusive spatial relationship data are consolidated into an Excel document by amalgamating the ‘Detected Class’, ‘Connection Information’, and ‘Crawling Data’ during the ‘Setexcelfiles.cs’ phase. The progression within ‘Setexcelfiles.cs’ is segmented into ‘Sort by Excel Data’, ‘Merge Data’, and ‘Write Excel’ stages, as visualized in Figure 6. In the ‘Merge Data’ step, the total area is calculated by adding the areas of the detected spaces and merging them with the sorted data. The ‘Write Excel’ phase culminates in the export of the spatial relationship data into an Excel sheet.

4.2. Experiment

4.2.1. Creating Training Data Using an Annotation Tool

Recognition involves annotating image regions using words [26]. Data annotation is defined as the labeling or tagging of relevant information in a dataset to allow machines to understand what they are [27]. It is difficult to extract data from raster-based floor plan images using polygonal methods. Given the variety of symbols and potential overlaps within floor plans, isolating specific data remains a complex task. This study facilitated the development of an annotation tool tailored to crafting training datasets from raster-based floor plan visuals. As illustrated in Figure 7, this tool streamlines the creation of labeled data by allowing users to demarcate classes using a bounding box. As discussed in Section 3.3, 21 label elements can be marked, encompassing both spaces and central classes. The periphery of the bounding box for the space class also includes a wall outline.

4.2.2. Data Training and Testing

The datasets for training and testing comprised 5524 and 1380 samples, respectively, which were split in an 8:2 ratio. The parameters designated for training encompassed a batch size of 64, a subdivision of 16, dimensions of 416 × 416 for height and width, and 78 filters. Test accuracy was ascertained using a confusion matrix [28,29]. This matrix consists of two rows and two columns delineating the true positives, false positives, false negatives, and true negatives [29]. When the actual true matches the predicted true, it becomes a true positive (TP). When the actual false is matched with the predicted true, it becomes a false positive (FP). When the actual true is matched with the predicted false, it becomes a false negative (FN). When an actual false result matches a predicted false result, it becomes a true negative (TN). The unique truth count (UTC) from the 1380 test data points was 52,253. Precision is calculated in Equation (1) as the ratio of how many positive predictions were correct (positive predictive value), as follows:

$$Precision={\displaystyle \frac{true positives}{predicted positives}}={\displaystyle \frac{TP}{TP+FP}}$$

(1)

The evaluation metrics for the space classes are listed in

Table 7. Precision for space classes.

Space Class	TP	FP	Precision (%)
Living room	916	49	0.949
Dining room	191	45	0.809
Bedroom	2998	118	0.962
Dressing room	1279	108	0.922
Closet	2986	434	0.873
Kitchen	894	27	0.971
Utility room	693	259	0.728
Pantry room	360	122	0.747
Laundry room	728	279	0.723
Bathroom	2196	106	0.954
Entrance	391	95	0.805
Stair room	1000	48	0.954
Stair entrance	884	121	0.88
Hallway	2282	379	0.858
Porch	928	156	0.856
Garage	802	13	0.984
Total	19,528	2359	0.873

. The TP and FP values were 19,528 and 2359, respectively, with an average precision of 0.873. The kitchen had the highest precision at 0.971, followed by the bedroom (0.962), stairs room (0.954), and living room (0.949). By contrast, spaces such as laundry rooms (0.723), utility rooms (0.728), and pantry rooms (0.747), which are intrinsic to households but lack distinguishing features, registered lower precision scores. The kitchen had high precision because the detected classes were small and the number of incorrect answers was small.

Table 8. Precision for central classes.

Central Class	TP	FP	Precision (%)
Single door	10,748	474	0.958
Double door	1228	105	0.921
Sliding door	1876	341	0.846
Window	11,322	247	0.901
Opening	1933	406	0.826
Total	27,107	2573	0.89

presents the evaluation metrics for the central class. Here, the TP and FP were 27,107 and 2573, respectively, with an average precision of 0.89. Single doors had precision scores of 0.958, followed by double doors (0.921) and windows (0.901). These central classes demonstrated high precision owing to their distinct features and the prevalence of their detection. Sliding doors were occasionally misclassified as windows. The lower precision of the opening classes can be explained by the lack of distinct features. Of the 52,253 UTCs, 686 were not detected.

Recall is the percentage of correct predictions of actual positives (true positive rate), calculated using Equation (2). The TP was 46,635, the FN was 5618, and the recall was 0.89.

$$Recall\left(Sensitivity\right)={\displaystyle \frac{true positives}{all positives}}={\displaystyle \frac{TP}{TP+FN}}$$

(2)

The bounding box detection rate is given by Equation (3): the TP was 46,635, the FP was 4932, and the UTC was 52,253, respectively, resulting in a detection rate of 0.99. This signifies that even when classes were not precisely matched within the bounding box, 99% of the detections were feasible.

$$Detection rate={\displaystyle \frac{actual detected value}{unique truth count}}={\displaystyle \frac{TP+FP}{UTC}}$$

(3)

The performance of the algorithm for object detection was evaluated using the mAP, as shown in Equation (4). Following the establishment of an IoU threshold of 0.5, the test yielded an mAP of 0.85. This score surpasses those documented in prior studies [11,18,30]. The results of the algorithm’s performance for data testing are shown in Figure 8.

$$mAP={\displaystyle \frac{1}{N}}{\sum }_{i=1}^N{AP}_i$$

(4)

4.3. Result

4.3.1. Results of Spatial Relationship Diagram Extraction

Upon the input of a floor plan image, the YOLO model discerns spaces and central classes and extracts diagrams showing spatial relationships. The detection rate is 99%. With an mAP of 85%, this method marginally outperformed previous studies [11,18,30] in the recognition of floor plan images, thereby affirming the suitability of the YOLO model parameters utilized in this test. Although discerning boundaries in completely open direct-connection spaces posed a challenge, it was reduced through the incorporation of an open class in the learning phase. While the stairroom was identifiable, establishing connections with the adjacent spaces remained problematic. This issue was addressed by introducing a stair entrance class in the learning phase to facilitate connections with neighboring spaces. In the future, the accuracy and performance of the model can be improved by training on many floor plan images. Figure 9 shows the result of spatial relationship diagram extraction from the program ‘AIBIM_Bubblemaker [19]’ developed in this study.

4.3.2. Results of Spatial Relationship Data Extraction

Table 9. Results of spatial relationship data extraction (part of Appendix B).

Project and Space Names						Space Size							Doors and Windows								Adjacent Spaces
						Ratio		Length Ratio		Area (sq/ft)	Space Area Ratio (%)	Building Area (sq/ft)	Class12	Number	Class13	Number	Class14	Number	Class15	Number	Space Name	Class Number	Random N	Wall	Door	Weight (Conversion)
Project ID	Layer	Floor Plan ID	Space Name	Class Number	Random N	Horizontal (%)	Vertical (%)	Horizontal (Ratio)	Vertical (Ratio)
20-573	1	20-573-1	Dining room	20	1	0.1748	0.2998	0.9701	1	96	0.0524	1835	0	0	0	0	0	0	1	1	Entrance	0	1	1	1	0.5

presents the results of the vector data extraction from the raster-based floor plan images. Further details of Table 9 are provided in Appendix B. The algorithm effectively extracts spatial relationship data and translates them into consistent vector data. Conceptually, a diagram can be interpreted as a graph with defined nodes and edges; this information is structured as an adjacency matrix. Four factors contributed to the extracted vector data, as follows:

• Project and Space Names: These were automatically sourced and include the project ID, layer, detected space name, class number, and a random “N” when there are identical space names;
• Space Size: This is determined as the percentage of the detected space area. For a total image size scaled to 100%, both the horizontal and vertical ratios were calculated. In raster images, the relative size of a space is typically inferred from the size of the door. The length of the side toward the door was standardized to 1 and the opposite side was scaled proportionally. The actual space area was derived by multiplying the total area (determined through data crawling) by the space ratio;
• Doors and Windows: These features are coded as binary data, where 1 indicates presence within the detected space and 0 indicates absence. For example, in the detected space, a single door (class12), double door (class13), sliding door (class14), and window (class15) were designated as either 1 or 0. The numbers of corresponding classes were also identified;
• Adjacent Spaces: These data were structured as an adjacency matrix, highlighting the connections of a particular space. If there was a wall or door connecting the space, it was marked as 1; otherwise, it was marked as 0. As per the adjacency definitions in Section 3.3, the connection weights for Cases 1, 2, and 3 were automatically set to 1, 0.5 for Case 2, and 0.25 for Case 3.

Upon comparing the outputs of the algorithm with real measurements, a minor discrepancy of 0.6% in the space size was noted. This variance can be attributed to the fact that the actual space measurements considered the inline of the wall, whereas the detected space accounted for its outline. Nonetheless, other vector information for the detected space remains in a consistent format. This consistency confirmed the accuracy and dependability of the proposed algorithm.

5. Discussion

5.1. Comparison of Dataset-Related Research

From an architectural perspective, obtaining raw data poses challenges owing to privacy constraints and copyright issues. As shown in

Table 10. Comparison of dataset-related research.

Dataset Name (Year)	Public Access	Model	Annotation	Number of Plans	Performance	Applied Studies
CubiCasa5K (2019)	○	Multi-task Model	80 object categories (doors, windows, walls, etc.)	5000	Rooms (Mean accuracy 69.8%) (MIoU 57.5%)	Indoor elements classification [23], Vectorized 3D reconstruction [31], Spatial data recovery [32], Floor plan layout analysis [33]
Korea LH (2019)	○	DeepLabV3+	5 elements (walls, windows, hinged doors, sliding doors and evacuation doors) and 7 space (rooms, entrances, balconies, dress rooms, bathrooms, living rooms, evacuation space, and pantries)	343	Elements (MIoU 80.68%) Room (MIoU 81.75%)	Generating GAN-based zoning [18]
LIFULL (2020)	X	-	dataset contains the data of LIFULL HOME’S, a Real Estate Information Service in Japan	5,300,000+	-	Apartment structure estimation [34], Application to similar property retrieval [35,36]
RFP (2021)	○	YOLOv4, DeepLabv3+	4 elements (wall, window, door, doorway) and 7 room types (balcony, bedroom, kitchen, other, library, living room, toilet)	7000	Elements and Rooms (MIoU 85%)	-
AIBIM_House (2023)	○	YOLOv3	16 space class(living room, dining room, bedroom, etc.) and 5 medium class (doors, windows, etc.)	14,970	Space classes (mAP 87.3%) Mediums classes (mAP 89%)	Floor plan recommendation system [6]

, the number of public plans is very small. This limitation inevitably restricts the application of AI technology in the field of architecture. CubiCasa5K [13] has many annotated items but the number of datasets is small and its performance quality is low. Korea LH [18] experimented with apartment floor plan images; however, the number of datasets was very small and the performance quality was low. LIFULL [31] provides real estate information services for Japan and primarily uses raster-based raw data. RFP [30] uses YOLOv4 and DeepLabv3+ models to recognize elements and rooms; however, there are fewer annotation items, making it unlikely to correctly judge the room category in open spaces.

The burgeoning growth of image recognition techniques has spurred interest in the conversion of raster-based floor plans into vector data in the architecture. However, these raster images vary in detail and quality and their annotation criteria and extent differ. Additionally, due to the overlap of complex elements and symbols, the performance of deep-learning models is lower than in general fields. In previous studies, the extraction of vectorized spatial data from raster-based images was limited. This study illustrates the feasibility of deriving spatial relationship diagrams from raster-based floor plans, coupled with the consistent extraction of vector-based spatial data through a tailored algorithm. Testing using the YOLO model revealed a detection rate of 99%, an mAP of 85%, and an MIoU of 74.2%. The efficacy of the algorithm marginally surpasses that of previous studies.

5.2. Implication

If training datasets are constructed and learned using deep learning, new ideas can be generated using AI. The datasets considered in this study were utilized to develop a floor plan recommendation system [6]. This recommendation system developed a deep neural network approach using SimGNN and shallow networks with the teacher–student learning to quickly and accurately compute graph similarity, as measured by the graph edit distance, during the search operation. This approach enhances exploratory creativity in design. The automation of exploratory creativity can be extended to transformational creativity [37]. If diagrams and vectorized spatial relationship data are trained using a GAN or a GNN, AI can generate new diagrams. Moreover, if AI learns not only spatial relationships but also information about location, direction, and entities, it can be extended to automated research for spatial layout generation and recommendations in pre-design. The construction of such high-quality large-scale training datasets increases the possibility of design automation in the sense of strong AI.

6. Conclusions

The aim of this study was to develop a robust training dataset for AI-driven architectural spatial layout generation and recommendation. The dataset includes spatial relationship diagrams extracted from raster-based floor plan images using the YOLOv3 model, as well as vector-based spatial relationship data obtained through a specialized algorithm. This methodology enables consistent and reliable dataset construction, distinguishing it from previous approaches. Data testing results demonstrated a detection rate of 99%, with an mAP of 85% and an MIoU of 74.2%. The efficient collection of raw data through crawling facilitates rapid expansion to include diverse building data. Additionally, incorporating advanced labeling methods, such as semantic segmentation or polygons, could further enhance the granularity of the dataset.

Furthermore, this dataset lays a strong foundation for future research in AI-driven architectural design. We plan to extend our work by leveraging this dataset to train GAN to represent spatial relationships in graph form, which could facilitate the automatic generation of spatial layouts. Additionally, by incorporating GNN, we aim to develop a system that recommends floor plans based on similar spatial relationships. This dataset could also potentially be utilized in the development and training of architectural Large Language Models (LMMs), which would significantly contribute to advancements in AI-driven architectural design.