Jul 5, 2026

Code

Synthetic Scan Dataset Generation for Non-Static Objects

Abstract

We present a virtual 3D scanning framework developed in Unity for generating realistic synthetic datasets tailored to geometric reconstruction and completion tasks in the architecture, engineering, and construction (AEC) domain. While deep learning approaches for scene and object completion have advanced rapidly, acquiring large-scale, high-quality real-world scan data remains costly and time-intensive. Unlike conventional synthetic datasets created through direct surface sampling or repositories such as ShapeNet, our approach simulates physically plausible scanning processes from virtual viewpoints, faithfully modelling sensor limitations, noise characteristics, and occlusions.

Workflow