Dimensionality reduction for enhanced 3D face recognition

This paper presents a novel approach for improving the accuracy of existing 3D face recognition algorithms via the dimensionality reduction of the feature space. In particular, two feature selection methods based on information criteria are selected and benchmarked herein (i.e. the minimum Redundancy - Maximum Relevance (mRMR) and the Conditional Mutual Information with Nearest Neighbors estimate (CMINN)), on top of the geometric features provided by a state-of-the-art 3D face recognition algorithm. Experimental validation on a proprietary dataset of 53 subjects illustrates significant advances in performance of the proposed method when compared to the reference 3D face recognition system. The repeated computations on several non-overlapping, randomly selected, training and test sets from the ensemble of frames, give evidence for successful classification of the subjects based on a significantly reduced subset of features with smaller cardinality, as obtained by CMINN. Finally, the high recognition capacity of this small fraction of biometric features is validated by the convergence of both methods to the same level of classification accuracy as the size of the utilized feature subset increases.