A Step from Virtual to Reality: Investigating the Potential of a Diffusion-Based Pipeline for Enhancing the Realism in Fully-Annotated Synthetic Construction Imagery

Abstract

In the rapidly evolving field of artificial intelligence (AI), synthetic data generation has become increasingly crucial, particularly in domains where real-world data is scarce, expensive, or sensitive. In this study, we introduce BCGen, a novel image realism enhancement pipeline that integrates our proprietary synthetic construction data generation and autonomous labeling engine, BlendCon, integrated with Generative AI. Leveraging the graphical capabilities of Blender and the deep learning prowess of the ControlNet model, BCGen represents a novel approach to synthesizing and enhancing construction site imagery. Our methodology narrows the reality gap, delivering images with increased realism and diversity while preserving the full annotations. The paper delineates our approach, methodology, and the broader implications of our findings. Through meticulous hyperparameter tuning and an innovative post-processing technique, we demonstrate the enhanced realism and diversity of the generated images, pointing towards the vast potential of synthetic data in visual AI applications within construction.