Deep Learning for Luminaire Object Detection in Floorplans

We present a deep neural network (DNN) machine learning model to detect luminaire objects from custom floorplan images, which contain small luminaire objects, complex backgrounds, and varying degrees of occlusion. We explore various training strategies, DNN architecture designs, and optimization methods in the literature. We custom train and optimize these and experimentally show that YOLO-v8m gives the best performance-accuracy trade-off, achieving a mean average precision of 0.48, perfectly within acceptable limits of inference time on a standard desktop computer. This represents a substantial decrease in labor and costs compared to manually annotating the floorplans, which takes around 45 minutes for an average size floorplan for highly trained personnel.