Inferring personal daylighting preferences using HDRI and deep learning techniques

Abstract

This paper presents a new method for classifying daylighting preferences based on deep learning models trained with pixelwise similarity features extracted from pairs of luminance maps. A new composite luminance similarity index is developed, which utilizes the pixelwise information from the entire luminance distribution and considers both the direction and magnitude of relative luminance change, instead of instantaneous metrics used in previous studies. The generated luminance and contrast similarity maps were directly used for training convolutional neural network (CNN) models to classify the occupant's visual preferences. Daylighting preference datasets for 11 individuals, collected in real offices using pairwise HDR images with simultaneous binary preference feedback, were used to evaluate the preference classification performance. The results proved the superiority of the luminance similarity index map as a preference indicator variable. CNN models trained with luminance similarity index maps showed impressive classification accuracy for all tested subjects in the dataset and exhibit more stable training and higher test accuracy compared to models trained with raw luminance maps. Common static lighting parameters cannot estimate daylight preferences even when used in powerful computational models; they neglect visual information located in various parts of the visual scene and cannot consider the change in perceived luminance distribution. Utilizing the full potential of HDRI sensing through detailed luminance mapping and feature extraction is an important step toward human-centered daylighting operation.