Data-intensive image based relighting

Image based Relighting(IBRL) has attracted a lot of interest in the computer graphics research, gaming, and virtual cinematography communities for its ability to relight objects or scenes, from novel illuminations captured in natural or synthetic environments. However, the advantages of an image-based framework conflicts with a drastic increase in the storage caused by the huge number of reference images pre-captured under various illumination conditions. To perform fast relighting, while maintaining the visual fidelity, one needs to preprocess this huge data into an appropriate model. In this paper, we propose a novel and efficient two-stage relighting algorithm which creates a compact representation of the huge IBRL dataset and facilitates fast relighting. In the first stage, using Singular Value Decomposition, a set of eigen image bases and relighting coefficients are computed. In the second stage, and in contrast to prior methods, the correlation among the relighting coefficients is harnessed using Spherical Harmonics. The proposed method thus has lower memory and computational requirements. We demonstrate our results qualitatively and quantitatively with new generated image data.