Exploring Reliability for Automatic Identity Verification with Statistical Shape Models
Federico Mateo Sukno, Alejandro Federico Frangi · 2007
Abstract way to detect the failure: it does not provide any reliability measure of the result. In fact, the segmentation process is One of the drawbacks ofstatistical shape models is their usually stopped due to the number of iterations rather than occasional failure to converge. Although visually this fact a convergence criterion. This situation changes in AAMs, is usually easy to assess there is not, in general, an auto- since they seek for the convergence of the represented tex-matic way to detect the failure. In this work we introduce a ture, but they can still converge to a wrong result [4]. generic reliability measure for statistical shape models. It This problem is especially important when ASMs are in-is based on a probabilisticframework and uses information tended to be used into fully automatic systems. For exam-extracted by the model itself during the matching process. ple, face recognition applications such as [10, 8] rely on the The proposed method was validated with two variants of segmentation of facial features to determine identity. A re-Active Shape Models by working with facial images. The sult like the one showed in Fig. 1 hampers any recognition experimental results showed a high correlation between the attempt. Avoiding (or at least automatically detecting) this segmentation accuracy and the estimated reliability. kind of situation may be of great advantage. Some (partial) solutions have been proposed in the liter-ature, mainly focused on specific applications. Wan et al.