samples with matching score below the updating thresh-old, as shown by the thinner edges (Fig. 5). However, the bottleneck represented by the edges related to sample localized in sub-graph G3 (belonging to J2) is discarded for adaptation.
Regular adaptation of the templates ensures its enhanced representation capability to the intra-class variations of the input samples. Coherently with this working assumption, we investigated if the updated templates resulted in an increase in the matching score values, on comparison with the test samples. This is done using the performance evaluation procedure described in Section 4.2 (Part C). In fact, increase in the matching score values means an increase in the similarity of enrolled templates to the test samples. Hence, an increase in the representativeness of the templates to the test samples acquired in different environmental conditions.
These results are reported in Table 5. For the baseline system, scores are computed using the two initial enrolled templates. It can be noticed that an average increment in the matching score is higher for graph min-cut in com-parison to self-update for all the traits. Thus, both the techniques ensure an increase in the template represen-tativeness, with the high superiority of graph min-cut over self-update.
To sum up, reported experimental results show that the behavior of self-update and graph min-cut is well described by the proposed conceptual representation.
These experiments also explain the reason of limited amount of intra-class variations captured by self-update, when a stringent threshold is used. At the same time, pointing out the ability of graph min-cut in localizing bottlenecks for labeling the input samples to be used for adaptation.
Experiments conducted so far have been done in order to validate the proposed conceptual representation and did not explicitly consider impostor samples. In the next section, we performed experiments under realistic condi-tions where the availability of impostor sample is quite probable during the system operation. To simulate the process of adaptation in real systems, batch of samples available for adaptation also includes impostor samples. 4.3.2. Evaluation of the self-update and graph min-cut techniques under realistic conditions Table 6 shows the performance of these updating schemes at different threshold values, that is, from 0.001% FAR to 1% FAR operational points. The performance is