Relighting the scene by finding the intrinsic images

Abstract

In this project we will embed virtual objects in a real-life scene looking auto- matically for the most realistic approach. Previous works have demonstrated the lighting as the most important feature to achieve a better and more realistic results. Moreover other researches that focus their works in which is the best way to descompose or classify the lighting of any scene. In this project, we have taken most of these concepts to begin our proposal. As a first step we didn't wish to get the real lights features or even their locations, our first goal was to relight the whole scene with a virtual light and do it in real-time. Proposal. Therefore, our goal is to be able to taking a frame from an input device (mostly Kinect), erase the current light conditions of the image to lately add new ones that must be common with the virtual objects included as augmented reality (AR). Create the new conditions will be done by setting up the virtual lights into a 3D enviroment and then use a render to extract the lighting and add it to the original image taken by the Kinect, all of it in real-time. The whole process itself, presents many challenges that must be faced. A way to isolate the lighting information from any image has to be find, the geometry of the scene or at least some significant areas must be retrieved to interact as similar as possible with the virtual lights and finally, a framework has to be developed to achieve the real-time objective for the whole task.