A Data-Driven Approach for Direct and Global Component Separation from a Single Image

The radiance captured by camera is often under influence of both direct and global illumination from complex environment. Though separating them is highly desired, existing methods require strict capture restriction such as modulated active light. Here, we propose the first method to infer both components from a single image without any hardware restriction. Our method is a novel generative adversarial network (GAN) based networks which imposes prior physics knowledge to force a physics plausible component separation. We also present the first component separation dataset which comprises of 100 scenes with their direct and global components. In the experiments, our method has achieved satisfactory performance on our own testing set and images in public dataset. Finally, we illustrate an interesting application of editing realistic images through the separated components.