Deep Light Direction Reconstruction from single RGB images

In augmented reality applications, consistent illumination between virtual and real objects is important for creating an immersive user experience. Consistent can be achieved by appropriate parameterisation of the model, that with real-world lighting conditions. this study, we developed a method to reconstruct general light direction from red-green-blue (RGB) images scenes using modified VGG-16 neural network. We reconstructed as azimuth elevation angles. To avoid inaccurate results caused coordinate uncertainty occurring at steep angles, further introduced stereographically projected coordinates. Unlike recent deep-learning-based approaches reconstructing source direction, our approach does not require depth information thus rely on special red-green-blue- (RGB-D) input.