Synthetic image generation for training 2d segmentation models at scale for computer vision progress monitoring in construction

Deep learning recognition models have been widely studied to recognize construction objects from site images. These methods require high volumes of quality data from human-made annotations to achieve moderate performance. Yet, manual annotation of images is time-consuming and error-prone, limiting ground-truth quality and performance of the resulting recognition models. To address such inefficiencies, automatic data generation and labeling of synthetic images have been recently explored. The quality of this type of data and how it can be effectively incorporated into a machine learning training pipeline still deserved further evaluation such that these models can scale up to real-world applications such as computer vision-based construction progress monitoring. This paper aims to re-evaluate synthetic image generation using physics-based simulation environments and 3D BIM. Using experimental results, insights are shared on how the quality of synthetic data can impact the performance of the trained recognition models where synthetic images with repetitive architectural or MEP patterns or 4D BIM with low-LOD engineering disciplines are used and when real and synthetic images are integrated into the same training pipeline. A path forward for improving the performance of the synthetic data and specifically mitigating the impacts of low LODs in BIM engineering discipline are discussed.