A major focus of research in computer graphics is the modeling and animation of realistic human faces. Modeling and animation of facial expressions is a very difficult task, requiring extensive manual manipulation by computer artists. Our primary hypothesis was that the use of machine learning techniques could reduce the manual labor by providing some automation to the process.
The goal of this dissertation was to determine the effectiveness of using an interactive genetic algorithm (IGA) to generate realistic variations in facial expressions. An IGA's effectiveness is measured by satisfaction with the end results, including acceptable levels of user fatigue. User fatigue was measured by the rate of successful convergence, defined as achieving a sufficient fitness level as determined by the user. Upon convergence, the solution with the highest fitness value was saved for later evaluation by participants with questionnaires. The participants also rated animations that were manually created by the user for comparison.
The animation of our IGA is performed by interpolating between successive face models, also known as blendshapes. The position of each blendshape's vertices is determined by a set of blendshape controls. Chromosomes map to animation sequences, where genes correspond to blendshapes. The manually created animations were also produced by manipulating the blendshape control values of successive blendshapes.
Due to user fatigue, IGAs typically use a small population with the user evaluating each individual. This is a serious limitation since there must be a sufficient number of building blocks in the initial population to converge to a good solution. One method that has been used to address this problem in the music domain is a surrogate fitness function, which serves as a filter to present a small subpopulation to the user for subjective evaluation. Our secondary hypothesis was that an IGA for the high-dimensional problem of facial animation would benefit from a large population made possible by using a neural network (NN) as a surrogate fitness function. The NN assigns a fitness value to every individual in the population, and the phenotypes of the highest rated individuals are presented to receive subjective fitness values from the user. This is a unique approach to the problem of automatic generation of facial animation.
Experiments were conducted for each of the six emotions, using the optimal parameters that had been discovered. The average convergence rate was 85%. The quality of the NNs showed evidence of a correlation to convergence rates as measured by the true positive and false positive rates. The animations with the highest subjective fitness from the final set of experiments were saved for participant evaluation. The participants gave the IGA animations an average credibility rating of 69% and the manual animations an average credibility rating of 65%. The participants preferred the IGA animations an average of 54% of the time to the manual animations. The results of these experiments indicated that an IGA is effective at generating realistic variations in facial expressions that are comparable to manually created ones. Moreover, experiments that varied population size indicated that a larger population results in a higher convergence rate.
| Identifer | oai:union.ndltd.org:nova.edu/oai:nsuworks.nova.edu:gscis_etd-1309 |
| Date | 01 January 2012 |
| Creators | Smith, Nancy T. |
| Publisher | NSUWorks |
| Source Sets | Nova Southeastern University |
| Detected Language | English |
| Type | text |
| Format | application/pdf |
| Source | CEC Theses and Dissertations |
Page generated in 0.002 seconds