Fig 1.
Features of the Pearson r correlation coefficient and the ICC.
(a) For illustration purposes only, we have drawn a random sample of 10 “subjects” from a population with mean = 100 and SD = 10. These are the Time1 data for all plots. In this case, we have set Time2 = Time1, and both the Pearson r and the ICC are equal to 1.0. (b) In this figure, the Time2 data are equal to the Time1 data but with added noise drawn from a population with mean = 0 and SD = 10. Both the Pearson r and the ICC decline somewhat, which is appropriate because the relationship between Time1 and Time2 is now weaker. (c) For the Time2 data for this plot, we did not add noise, but merely added 20.0 to the Time1 data. This is essentially representing a “session” or “occasion” effect. In this case, the ICC has declined substantially but the Pearson r is still at a perfect 1.0. (d) In this case, we did not add noise to Time1 data, but performed a linear transformation of Time1 data to get Time2 data. The ICC has dropped precipitously whereas the Pearson r is still at 1.0. This is what we mean when we say the that the Pearson r is invariant to linear transformation. The 4 variances below the ICC are used to calculate the ICC, as explained in the S1 Document.
Table 1.
Full names and abbreviations for eye movement databases.
Fig 2.
Analysis of the Synthetic database.
Top plot: ICC Histogram. Low ICC features (black bars), moderate ICC features (gray bars), high ICC features (white bars). Middle plot presents the Rank-1-IR results as a function of the number of PCA components. Lower plot presents the EER results. Each symbol represents the median over 100 random training-testing sets.
Table 2.
Description of databases.
Table 3.
List of features for CEM-12-DB, OPC-18 and CEM-14 databases.
Table 4.
Brain structures measured in the BRAIN database.
Fig 3.
Flowchart of the data analysis.
See text for details.
Table 5.
Biometric performance summary analysis.
Fig 4.
Analysis of the SBA-Short-Term database.
See caption for Fig 2.
Fig 5.
Analysis of the SBA-Long-Term database.
See caption for Fig 2.
Fig 6.
Analysis of the CEM-12-DB-Short-Term database.
See caption for Fig 3.
Fig 7.
Analysis of the CEM-14-Long-Term database.
See caption for Fig 2.
Fig 8.
Analysis of the OPC-18-Short-Term database.
See caption for Fig 2.
Fig 9.
Analysis of the OPC-18-Long-Term database.
See caption for Fig 2.
Fig 10.
Analysis of the CEM-14-Short-Term database.
PCA was not performed, due to the small number of features. All features were entered directly into the biometric assessment algorithm. Plotted are means (dots) and minimum and maximum value error bars based on 100 training and testing sample sets.
Fig 11.
Analysis of the CEM-14-Long-Term database.
See caption for Fig 10.
Fig 12.
Analysis of the GABM database.
See caption for Fig 2.
Fig 13.
Analysis of the GZVMB database.
See caption for Fig 2.
Fig 14.
Analysis of the FACE1 database.
See caption for Fig 2.
Fig 15.
Analysis of the FACE2 database.
See caption for Fig 2.
Fig 16.
Analysis of the BRAIN database.
See caption for Fig 2.
Fig 17.
Relationship between biometric performance and ICC across databases.
(A) Scatterplot relating Rank-1-IR and median ICC. Note the strong linear relationship. Also shown are the best fitting line, the best estimate of the linear equation, the results of an F-test, and an r2 estimate. (B) Scatterplot relating EER and median ICC. This relationship is even stronger than that for Rank-1-IR, with ICC accounting for 92% of the variance in EER.
Fig 18.
Relationship between the median of the similarity score distribution and ICC.
(A) Scatterplot relating the median of the similarity score distribution for the genuine scores and median ICC. Note the strong linear relationship. Also shown are the best fitting line, the best estimate of the linear equation, the results of an F-test, and an r2 estimate. (B) Scatterplot relating the median of the similarity score distribution for the impostor scores and median ICC. This relationship is weak and not statistically significant.
Fig 19.
Relationship between the IQR of the similarity score distribution and ICC.
(A) Scatterplot relating the IQR of the similarity score distribution for the genuine scores and median ICC. Note the statistically significant linear relationship. Also shown are the best fitting line, the best estimate of the linear equation, the results of an F-test, and an r2 estimate. (B) Scatterplot relating the IQR of the similarity score distribution for the impostor scores and median ICC. This relationship is weak and not statistically significant.
Table 6.
Published performance for other EM-related approaches.