Data-Efficient Scan-to-BIM: A Weakly Supervised Learning Framework for 3D Point Cloud Analysis in Modular Construction

This research proposal aims to automate the Scan-to-BIM Quality Control (QC) process for Modular Off-site Construction (MOC).
MOC requires millimeter-level precision, but the 3D scan data (point clouds) consist of raw coordinates without semantic information like wall or floor. Using AI (3D Semantic Segmentation) to automate classification presents a Catch-22 dilemma: training the AI requires manually labeled 3D data, which is prohibitively expensive and time-consuming.
This study proposes a Semi-Supervised Learning (SSL) and Self-Supervised Learning (SSL) framework to overcome this problem by drastically reducing the dependency on labeled data.
The core idea is to pre-train the AI model on large-scale unlabeled data, allowing it to learn the grammar of 3D structures on its own. Subsequently, a high-performance segmentation model is built using only a minimal amount of labeled data (approx. 1-5%).
The ultimate goal is to automate the Scan-to-BIM post-processing workflow, prevent costly on-site rework, and contribute to enhancing the MOC industry’s competitiveness.

Faculty Supervisor:

Sang Hyeok Han

Student:

Partner:

Gyeongsang National University

Discipline:

Engineering

Sector:

Construction; Technology

University:

Concordia University

Program: