| Title |
Establishment of a Spatial Configuration Interpretation Process for Architectural Floor Plans and Performance Evaluation Using General-purpose Large Multimodal Models and Domain-specific Vision Models for Architectural Drawings |
| Authors |
서예지(Seo, Yeji) ; 허민지(Heo, Minji) ; 구형모(Gu, Hyeongmo) ; 추승연(Choo, Seungyeon) |
| DOI |
https://doi.org/10.5659/JAIK.2026.42.7.127 |
| Keywords |
Artificial intelligence; LMM; Multimodal Models; Architectural Drawing Interpretation; Architectural design support; Floor plan recognition; bubble diagram. Floor Plan Interpretation Process |
| Abstract |
Recent advancements in artificial intelligence technology have highlighted the need to utilize it as a new design support tool to streamline the
architectural process. The architectural design process involves the use of diverse data formats, making the application of large multi-modal
models(LMM) capable of understanding multiple data types increasingly important. However, architectural floor plans contain abstract
information such as symbols and scales, presenting limitations for general-purpose LMM to interpret accurately. Therefore, this study aims to
expand the potential for AI utilization in the architectural design process by proposing a workflow that combines an architecture-specific
vision model with a general-purpose LMM. This study visualizes architectural object recognition data and adjacency relationship data from
floor plans in a bubble diagram format, extracting this information as JSON formatted data and natural language data. The extracted data is
used as input for the general-purpose LMM, and its inference performance in response to user queries is analyzed. Analysis results show that
the proposed workflow demonstrates superior accuracy and processing speed in architectural floor plan interpretation compared to using a
general-purpose LMM alone. This research is significant in that it shifts floor plan interpretation from an empirical reading approach to a
data-driven structural analysis approach. |