Shape Modeling and GPU Based Image Warping

Abstract

This project addresses the problems of manually placing facial landmarks on a portrait and finding a fast way to warp the annotated image of a face. While there are many approaches to automatically find facial landmarks, most of them provide insufficient results in uncontrolled environments. Thus I introduce a method to manually adjust a non-rigid shape on a portrait. This method utilizes a statistical shape model based on point distribution models. With these manually placed landmarks the image of a face can be warped into another shape. To warp the image I use a piecewise affine transformation. This way of transforming, however, tends to be computationally intense and therefore slow. Thus in the second part of the project I introduce a way to perform a piecewise affine transformation with enhanced performance using shaders in OpenGL. This project is made in collaboration with the Pedagogical University of Berne, Switzerland and will be part of a system for diversity research named chic-o-mat. Eventually the system will run on an iPhone as an application available to the public. Therefore, the provided solutions are based on iPhone programming using the multi-touch screen for the shape adjustment and the GPU of the latest iPhone 4S. A test application demonstrates up to 20X speedup performing piecewise warping using the GPU.