Realistic image compositions with deep learning priors

Image composition is an operation that incorporates an object extracted from a source image to incorporate it inside a target image. This simple operation has a lot of practical applications in the advertisement, entertainment, and movie industries. However, it is tedious for an artist to achieve realistic image compositions. Indeed, it requires different lighting, shadows, or objects present in the target image will change the target object's appearance. One of these appearance changes is new contact shadows between the target object and the target environment.

Spatially-aware lighting estimation

We present a method for automatically estimating the lighting conditions from a single image. As opposed to most previous works which proposed methods that estimate only global lighting or use a limited illumination representation (low frequency SSH, parametric model), the proposed method attempt to use a new spatially-varying light representation with realist texture to render shiny objects. Our method will estimate a coarse (cuboid) geometry of an indoor scene from a single image and use this geometric information to hallucinate a realistic room texture used for illumination.