DeepLux: A data-driven method and benchmark for 3D illuminance maps estimation

Abstract

In this paper, we propose a pipeline and benchmark, called DeepLux, for predicting illuminance on 3D point clouds. Classic algorithms for computing photometrically accurate illumination are based on numerical and analytical models which are generally computationally expensive, especially in scenarios with complex geometries. Unlike existing approaches, our algorithm is the first learning-based method that is able to predict accurate illuminance map information that could be used for real-time smart lighting applications. We also evaluate our approach on two complementary tasks, that is, light position and intensity estimation, which are important aspects in the field of lighting design. Additionally, we provide an extensive novel photometrically correct dataset, which we use for training and evaluating our approach. Experiments show that the proposed algorithm produces results on par with or even better than the state of the art (+8% average error in real tests) while achieving faster simulation timings than its analytical counterpart, especially in complex synthetic and real-world scenarios.