Classification and Performance Evalua...

	In this task we want to investigate the usefulness of several classification approaches with the single purpose of recognize the identity of individual just by analyzing the temporal dynamics of faces in 3D when expressions or speech are performed by the individual.  [Demos] [Publications]

T4 - Classification and Performance Evalua...

In this task we want to investigate the usefulness of several classification approaches with the single purpose of recognize the identity of individual just by analyzing the temporal dynamics of faces in 3D when expressions or speech are performed by the individual. We want to follow two lines of research: First we want to investigate the usefulness of the facial dynamics in the process of identity recognition; simultaneously, we want to investigate the possibility of simultaneously recognize the identity of the individual and classify the type of expression he is presenting. All of this must be accomplished in 3D and should be head pose and illumination invariante. Due to the type of face modeling that will be used (AAM), this research will address in depth the problem of partial face occlusion. In this task, the classification issues we propose to investigate are mainly pattern recognition tools. The typical paradigm for evaluating classification algorithms is to divide a data resource into training and test collections, or to use cross-validation if the resource is relatively sparse. Therefore, it is important for acquire a representative facial motion data collection, that will be held in consideration in task 2.

- Load training and testing feature data from the previous tasks The extracted feature data is loaded from task 3 that holds “high-level” training and testing data from the experiences performed earlier in task 2. All sub-datasets (each experience) holds both the features described the using facial muscle activation based models and in low dimensional manifolds.

- Classification using current state of the art methods This task consists in apply the current state of the art methods for classification, namely Support Vector Machines (SVM) and Hidden Markov Models (HMM) , oriented to multiples classes, to the previous extracted features. These will be used to derive matching or likelihood scores for determining identity and expression of the testing subject.

- Evaluation and performance metrics The evaluation of the overall system will be provided by the standards performance metrics, in particular: False Reject Error, False Acceptance Rate, True Acceptance Race, Receiver Operating Characteristics (ROC), confusion matrices, etc.