LightNet: Deep Learning Based Illumination Estimation from Virtual Images

Abstract

In the era of virtual reality (VR), estimating illumination with lighting direction and lighting virtual objects has been a challenging problem. In VR, poor estimation of illumination and lighting direction makes any virtual objects into unrealistic. The inaccurate estimation of lighting can also cause strong artifacts in relighting of the virtual images. Inspired by these issues, the main objective of this paper is to enrich visual rationality of single image by providing accurate assessments of real illumination and lighting direction. We proposed a LightNet architecture by modelling Denseset121 network to estimate the light direction and color temperature level in any virtual reality images. We present quantitative results on VIDIT dataset to evaluate the performance and achieved good results in all the performance metrics. The experimental results proved that the proposed model is robust and provides a good level of accuracy in estimating illumination and lighting direction.