PDF version of this entire document

Technical Report

Roy Schestowitz

Abstract:

Statistics of faces are intricate and delicate when posed as a task requiring the ability to discern identities. In 3-D, where the data is essentially range images void of any textural information, the task is further complicated, especially in the presence of facial expressions. This work explores our ability to carry out verification tasks, where various tools are tested and compared, ranging from intensity-based PCA (superimposed upon range images) to more sophisticated algorithms such as Generalised Multi-Dimensional Scaling (GMDS). We find that the performance reaches its peaks at around 97% for particular datasets and hovers at the range of 90-95% for FRGC data, depending on the method being assessed. The work is unique in the sense that it thoroughly but not exhaustively explores what can be attained without any textural information, using tools which are claimed to have already been outperformed, e.g. by Al-Osaimi et al. We are unable to replicate results of such groups due to the proprietary nature of their data and methods, but by bringing together under the same roof several families of methods we make reasonably valuable benchmarks possible.