Results

FIW supports laboratory style evaluation protocols, Kinship Verification (i.e.,  one-to-one) and  Family Classification (i.e., one-to-many).

Kinship Verification

Kinship verification tests were run on previously tested pairs such as siblings, mother–daughter, father–son, etc. along with four newly introduced multi generational pairs between grandparents and grandchildren in the FIW database. With the Fine-Tuned CNN model, we saw an average of a 4% increase in accuracy from the VGG-Face model. The benchmark tests were completed by randomly selecting an equal number of positive and negative pairs for each family, then computing cosine similarity for each pair.  Picture
Averaged verification accuracy scores (%) for 5-fold experiment on FIW. Note that there was no family overlap between folds.

Family Classification

By fine-tuning the VGG-Face model, we achieved an average of a 1% increase in accuracy. This was accomplished by following the 5-fold cross-validation protocol with no family overlap on families with more than five members. Folds were then made up of one member for each family and multi-class SVM was used to model VGG-Face features for each family. Picture
Family classification accuracy scores (%) using 524 families.