Deep Learning for Luminaire Detection in Professional Lighting Images

This study evaluates vision-based deep learning models for real-time artificial light source detection on professional lighting images. We investigate detecting luminaires on (possibly embedded) edge devices and subsequently classifying them. The main challenges to overcome are the rather strict computational limitations of these devices, as well as the issues arising from image occlusion and reflections. We explore and compare various model training strategies, neural network architecture designs, and optimization methods. In our experiments, YOLOv9t achieves a mean average precision (mAP) of 0.65 at 86 fps, a 54.76% performance increase compared to feature-based model performance on edge cases.