Bridging 2D drawings and ai: evaluating the requirements and feasibility of machine learning models for quantity take-off and general interpretation of issued-for-construction drawings

Fu, Junryu

Permalink

https://hdl.handle.net/2142/129739 Copy

Description

Title

Bridging 2D drawings and ai: evaluating the requirements and feasibility of machine learning models for quantity take-off and general interpretation of issued-for-construction drawings

Author(s)

Fu, Junryu

Issue Date

2025-05-08

Director of Research (if dissertation) or Advisor (if thesis)

Golparvar-Fard, Mani

Department of Study

Civil & Environmental Eng

Discipline

Civil Engineering

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Construction Management
• Quantity Take Off
• 2D Drawings
• Computer Vision
• Machine Learning

Abstract

This thesis aims to explore the feasibility of using Artificial Intelligence (AI) methods to automatically learn and interpret 2D drawings in the context of the built environment. 2D drawings—more specifically, Issued for Construction (IFC) drawings—are fundamental across all architecture, engineering, construction, and operation (AEC/O) use cases, where professionals often spend 70–80% of their time reading, analyzing, and generating such documents. IFC drawings are inherently complex, even for humans, as they consist of abstract visual representations combined with unstructured textual and visual information. Furthermore, these drawings often undergo multiple revisions even after the design phase due to unforeseen site conditions, shifting requirements, and coordination issues among various trades and stakeholders. As a result, manual and often repetitive work related to 2D drawings persists throughout the life cycle of a project, introducing inefficiencies and contributing to confusion in the process. To address these challenges, this thesis investigates the application of modern computer vision techniques to IFC drawings for tasks such as Quantity Take-Offs (QTO) and domain-specific interpretation. It evaluates the effectiveness of existing Machine Learning (ML) models—considering both architectural design and training data—that are typically developed using datasets not tailored to the AEC/O domain. The study explores the true capabilities, strengths, and limitations of these models within this context. In addition, it identifies key pain points and adoption barriers, highlighting the critical and largely unmet need for domain-specific datasets to advance future research and development. Building on these insights, the thesis introduces newly developed datasets to establish preliminary benchmarks and assess the performance of current models in interpreting complex 2D drawings. It also presents incremental improvements in algorithmic techniques and model comprehension of domain-specific drawings aimed at enhancing existing industry workflows. Through a series of experiments conducted on real-world IFC drawings, the limitations and potential of these approaches are analyzed in depth. Notably, even a modest 0.1% improvement in workflow efficiency could yield annual savings of $1.4 billion—demonstrating the substantial impact that AI models tailored to 2D drawing interpretation could offer. The thesis concludes with a detailed discussion of future research directions.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129739

Copyright and License Information

Copyright 2025 Junryu Fu

Owning Collections