O M 4 . 1 The Subject Database 4 . 2 Experiment Plan 5 . 1 Varying the Overlap 4 Experimental Setup 5 Parameterisation Results

Stochastic modelling of non-stationary vector timeseries based on HMMs has been very successful for speech applications [5]. Recently it has been applied to a range of image recognition problems [7, 9]. Previously reported work [6] has investigated the use of HMMs to model human faces for identi cation purposes. Faces can be intuitively divided into regions such as the mouth, eyes, nose, etc., and these regions can be associated with the states of an HMM. The identi cation performance of a top-bottom HMM compares favourably with some of the well-known algorithms, for example eigenfaces as detailed in [8]. However, the HMM parameterisation in the work presented so far was arrived at by subjective intuition. This paper presents experimental results which show how identi cation rates vary with HMM parameters, and which indicate the most sensible choice of parameters. The paper is organised as follows: section 2 gives an overview of the HMM-based approach; section 3 details the training and recognition processes; section 4 describes the experimental setup; section 5 presents the identi cation results; section 6 concludes the paper.