PDF version of this entire document

EDM Versus Pure Residual Approach: Correlation Between PCA-based Approaches

In page 13 of the IJCV paper there is a head-to-head comparison between GMDS- and EDMs-based similarity - one that takes what's said to be verification rate of 77% at $10^{-3}$ false acceptance rate (FAR). Their recognition rate is not too high, about 93%. The subsequent work involves manageable experiments as measured in terms of scale, using lab-collected data to construct person-specific EDMs and then plot recognition rates as ROC curves. Given GIP data from different subjects, this is reproducible.

Working with groups of similar expressions from the same subject (screenshots available), e.g. by taking residuals computed from

'~/Facial-Expressions-Recognition/fear.v3r'

to

'~/Facial-Expressions-Recognition/Suprise.v3r'

and then also considering the diffs from

'~/Facial-Expressions-Recognition/Suprise.v3r'

to

'~/Facial-Expressions-Recognition/Joy.v3r'),

we do have agreement between match score for the model-based approach which yields (for the first six pairings):

$1.0e-04*[0.0839,0.1897,0.1435,0.1322,0.1971,0.1946]$

And for the non-model-based similarity measure (mean of differences in the residuals):

$[0.0107,0.0190,0.0169,0.2056,0.2825,0.2562]$

The best match so far is image #1, based on both measures that agree (the model based and pure residual/difference-based).