3D facial reconstruction from 2D portrait imagery

Abstract

Whilst Facial Reconstruction is an intensely studied topic, it is still a highly debated topic on how to transition from a 2D portrait image of a person to a 3D facial model in an efficient and computationally acceptable manner. Despite being an ill-posed problem, the generation of a 3D facial model from a 2D image containing a human face is possible using proven methods and techniques. The aim of this project is to put forward an approach that methodologically extracts 3D human faces from 2D portrait images of people. This system will allow for the generation of 3D facial models based on the different facial poses that humans normally project. In addition to this, it allows for the extraction of the texture from the input image itself. By building a predictive regression tree model that can forecast the general outline of the face, mappings can be set between each of the key facial features to specific vertices on a reference 3D model. Subsequently, through the use of the concept of 3D Morphable Models, the necessary transformations are applied to the reference model to create the finalised and updated 3D face model alongside the 2D texture image. From the results acquired through the evaluation process it is shown that when processing various models generated from different regression trees using the Root Mean Square, 75th Percentile and Arithmetic Mean equations, similarities between models that are of the same person are identifiable. Specifically results showed the lowest accuracy and precision percentages were of 75% and 50% respectively, while the highest accuracy and precision rates were of 97% and 94% respectively. Through this evaluation, it is noted that the accuracy of the comparison values is dependent on the regression trees used for the alignment of the landmarks on the input image rather than the structure of the landmarks themselves.