건축 분야의 인공지능 활용을 위한 시각 데이터셋의 구조적 유형 분석에 관한 연구 = A Study on the Typological Analysis of Visual Datasets for Artificial Intelligence Research in Architecture

다국어 초록 (Multilingual Abstract)

In recent years, artificial intelligence (AI) has emerged as a key technology driving innovation across diverse industries, including architecture; however, its practical application in the architectural domain remains limited. One of the main causes is the lack of high-quality training datasets and insufficient quality management. The performance of AI models depends heavily on the scale, precision, and annotation accuracy of datasets, yet architectural data are difficult to publish or share due to copyright restrictions and domain-specific complexity, often existing in irregular formats. Accordingly, this study aims to systematically collect, classify, and analyze visual datasets applicable to AI training and benchmarking in architecture, in order to clarify their structural typologies and practical characteristics.

Following the PRISMA 2020 guidelines, a systematic literature review was conducted on major academic publications and web resources from 2005 to 2025. Out of 778 identified studies, 53 datasets related to architectural visual data were selected and analyzed using standardized indicators such as modality (photo, drawing, 3D), annotation type, data scale, accessibility, and application domain. As a result, architectural visual datasets were categorized into three primary modalities and eight subtypes. Photo-based datasets were divided into single-view, panoramic, and aerial imagery; drawing-based datasets were grouped into raster and vector formats; and 3D datasets were classified into depth-based views, geometric models, and attribute-based models.

By type, photo datasets are primarily used for object detection and style classification tasks, while drawing datasets serve as key resources for floor-plan recognition, generation, and object detection. Three-dimensional datasets are widely used for spatial understanding, simulation, and performance evaluation but face limitations in accessibility and standardization. Among these, BIM/IFC-based attribute models possess the highest potential but remain restricted by copyright and file-format compatibility issues.

This study moves beyond previous works that focused solely on individual dataset reports or case-specific utilization by providing a structural and quantitative comparison across dataset types. It proposes a standardized taxonomy and comparative indicators to help researchers and practitioners select datasets appropriate to their objectives. The findings contribute academic foundations for the systematic construction, sharing, and evaluation of architectural AI datasets and may serve as a reference framework for establishing a national-level architectural data hub and AI training infrastructure in the future.

목차 (Table of Contents)

• 제 1장 서론 01
• 1.1 연구의 배경 및 목적 02
• 1.2 연구의 대상 및 방법 05
• 제 2장 이론적 배경 및 선행 연구 분석 07
• 2.1 데이터셋의 유형화와 도메인 특수성 08
• 2.2 선행 연구 분석 12
• 2.2.1 인공지능의 성능과 데이터셋의 관계 12
• 2.2.2 건축 분야의 데이터셋 활용 및 구축 연구 16
• 제 3장 건축 분야의 시각 데이터 유형화 및 데이터셋 현황 18
• 3.1 건축 분야 시각 데이터 유형화 19
• 3.1.1 사진 유형 21
• 3.1.2 도면 유형 22
• 3.1.3 3D 유형 23
• 3.2 건축 분야 시각 데이터셋의 현황 조사 26
• 3.2.1 데이터셋 수집 및 선별 절차 26
• 3.2.2 데이터셋 구축 현황 29
• 제 4장 건축 분야 시각 데이터셋의 유형별 분석 및 비교 39
• 4.1 사진 기반 시각 데이터셋 분석 40
• 4.1.1 단일 시점 이미지 40
• 4.1.2 파노라마 이미지 43
• 4.1.3 원격/항공 이미지 46
• 4.2 도면 기반 공간 표현 시각 데이터셋 분석 49
• 4.2.1 래스터(Raster) 형식 도면 50
• 4.2.2 벡터(Vector) 형식 도면 52
• 4.3 3D 공간 구조 시각 데이터셋 분석 55
• 4.3.1 깊이 기반 시점 데이터 56
• 4.3.2 형상 모델 데이터 57
• 4.3.3 속성 기반 모델 데이터 60
• 4.4 유형별 비교 분석 및 활용 전략 63
• 4.4.1 유형별 데이터셋 특성 비교 분석 63
• 4.4.2 데이터셋 선택 시 고려 요소 도출 67
• 4.4.3 적용 사례 및 활용 시나리오 70
• 제 5장 결론 72
• 참고문헌 75
• 영문요약 86