Stereoscopic Pop-Up

I present a novel system for the segmentation of images with depth into a mesh-based layered representation. This layered representation can be used for synthesising new views of the scene, and to create effects which would otherwise be non-trivial to implement, such as camera pans and depth of field effects. My contributions are threefold. Firstly, my project builds on existing algorithms, in particular GrabCut (Rother et al., 2004) and exemplar-based inpainting (Criminisi et al., 2004), which I extended to use and generate depth information to improve their results. I discuss in detail the implementation of the various algorithms used and the extensions made, providing additional information on the implementations not covered in the original papers. Secondly, I introduce a novel method for automatically selecting foremost elements in an image, so that they can be extracted in front-to-back order. Thirdly, I describe a real-time rendering algorithm optimised for rendering images with depth, using the layered representation. I present several examples of the results of this project, including examples of the final result and intermediate steps.