Architectural plan computer-aided design (CAD) drawings constitute fundamental references for building whole life-cycle management. However, efficient querying, retrieval, and analysis of these drawings across departments are frequently hindered by persistent knowledge gaps among personnel, significantly undermining overall management efficacy. Visual question answering (VQA) technology, especially in the form of multimodal pretrained large models, has been widely adopted in image data management, significantly enhancing image analysis and retrieval capabilities. As specialized image data for professional use, architectural plan CAD drawings hold substantial potential for VQA-based analysis and retrieval. However, the absence of dedicated data sets and evaluation methods for semantic question and answering on architectural plan CAD drawings hinders the assessment and advancement of VQA applications in this domain. To address this gap, specialized data sets are urgently needed to support related research. To this end, this study first designs a structured question taxonomy for architectural plan CAD drawing VQA and establishes a multidimensional VQA database (ArchPlanVQA). Based on this framework, we further systematically evaluated the capabilities of current pretrained vision-language models (VLMs) in architectural spatial semantic parsing and engineering knowledge reasoning. The research reveals significant differences in color and texture characteristics, as well as semantic representations, between CAD-converted images and conventional daily life images; although the test shows that current general-purpose VLMs only achieve 33.03 to 37.88% semantic understanding accuracy. However, the evaluation results across different VLM architectures and frameworks demonstrate clear consistency, presented that the general-purpose VLMs have stable reasoning capabilities in semantic understanding tasks for architectural floor plan CADs. This paper provides foundational data and methodological frameworks to support and facilitate VQA technology in architectural engineering.
Li et al. (Tue,) studied this question.