Building Automation with Vision Transformer Using Synthetic Indoor Images for Room Light Control
Yuji Aizono, Takuya Sakamoto, Hideya Ochiai, Hiroshi Esaki · 2024
In recent years, there have been reports of using AI to automate building facilities along with the advancements of AI. In this context, we have introduced a new architecture that directly generates control signals for building facilities from room images using deep learning, distinguishing it from traditional object detection or reinforcement learning approaches. Although deep learning typically requires a dataset for training, it is often the case that there is no dataset available for the specific building when implementing a system, making dataset preparation a challenge during the initial stages of implementation. In this study, to address this issue, we proposed a method called Simple Indoor Image Synthesis (SIIS), a technique for synthesizing images, and conducted the evaluation of control accuracy using the synthesized images for model training. The evaluation results shows that higher accuracy, e.g., 98.4%, can be obtained with larger number of synthetic images, demonstrating that SIIS is practically useful for model training even at the initial stages of system implementation.