PDF version of this entire document

Introduction

Face identification (or verification) helps associate a person with his/her real identify when it cannot be reliably inferred from a document in one's possession. Authentication commonly requires that a person's biometric characteristics closely resemble a model or reference digitally stored - one corresponding to the same person. In order to reliably carry out this task, one can measure different attributes that are perceived to be immutable, e.g. resistant to illumination changes, pose-agnostic, and independent of the acquisition method. If for each person a large (but finite) set of surfaces can be considered a proper ``match'', then we wish to identify the subspace in which those surfaces lie. Only through good separation in a (very high) parametric space can one person never be confused with another. The problem is harder than typically realised, especially when depth information is all we are being provided. Since people's faces are topologically similar, it is the fine differences that tell apart one person from another and since the face is very morphologically flexible, there is no guarantee of a neutral expressions being presented all the time.

In Section 2, a very concise summary of past work is presented. In Section 3, a discussion of key concepts (and hypothesis) is outlined. Section 4 deals with data and a tiny portion of our practical work is presented in Section 5, emphasising only the best of our results. Section 6 concludes.