PDF version of this entire document

Early Results With Weighting

Taking the first imaged person (58 images) and enrolling it for an experiment (versus 92 random people), the ROC curve shows good performance. Some other people remain more problematic because they deliberately change expressions and complicate things around the eyes (further smoothing might be needed there), which leads to greater necessity for cunning weighting schemes. This will require some further exploration to work around.

Figure: ROC curve for the first phase of the experiment, which compares one-to-one (same person) and many-to-many (different people excluding this person, except in one case)

With borderline cases removed (detected and rejected), the results are without an error in detection. Without automated filtering of borderline cases (declination to classify), the ROC curve now looks as shown in Figure .

Figure: A broader scope curve for performance as in the previous figure

How many subjects get included may be important here. How many instances is important as well. The number of subjects is 82 for the false pairs and 5 for true pairs. I now work on expanding the latter.

Upon closer inspection of the problematic cases, improved initial alignment would help eliminate some of the borderline cases. Additional experiments have been run where increased emphasis is put in the nose for alignment, as the nose region is also weighted most heavily by the similarity measure.

Figure: An example of misalignment in some parts of the nose in a true pair of images (same person), with the left nostril being a prime example

Figure: An example of a borderline case (leaning towards false positive)

Figure: Debugging information for the problematic true pair shown before

Figure: Debugging information (distance differences) for the aforementioned false positive

Figure: The problematic (borderline) false positive after the new alignment scheme gets applied