Face recognition has been one of the most important tasks in computer vision due to the wide range of applications in several environments, such as surveillance systems, biometrics, forensics. One of the face recognition tasks, the face verification is responsible for  determining whether two facial images belong to the same subject (i.e., genuine matching) or are from different subjects (i.e., an impostor). An example of face verification is the following. Imagine you are at a bank ATM machine, the system will take your picture, extract its features and match them to your biometric information (template) stored in the bank database based on the account number you entered. If there is a positive match, you will have access granted to your account.

Face verification is the focus of our paper entitled Face Verification based on Relational Disparity Features and Partial Least Squares Models that will be presented in the Conference on Graphics, Patterns and Images (SIBGRAPI 2017).

Hypothesis i) two face images of the same subject would hold small differences, but this difference increases when we consider a pair of images from different subjects; ii) multiple classifiers might achieve higher verification rates due to higher diversity.

The next figure illustrates how two pairs of images are compared. Once their features are extracted using the VGG-Faces network, they are subtracted element-wise, resulting in the feature vector that will be presented to the classifiers. Such vector is referred to as disparity feature vector.

Approach two sets of disparity feature vectors (same and not same) are used to learn each of k Partial Least Squares (PLS) model. To increase the diversity, each PLS model is estimated using sets with different samples. Then, the disparity features are extracted from a pair of testing images to compose a feature vector which is classified by each PLS model, from which the response values are used to estimate the label (genuine or impostor), based on a majority voting scheme.

Note we used outside data (samples not belonging to the LFW dataset) only for the feature learning once we used the VGG-Faces network, but we have not used outside data for learning pairs same and not same as most of methods do. The LFW dataset should devise a protocol specific for such cases, otherwise the comparisons are not completly fair.


Rafael Henrique Vareto, Samira Santos da Silva, Filipe de Oliveira Costa, William Robson Schwartz. Face Verification based on Relational Disparity Features and Partial Least Squares Models. Conference on Graphics, Patterns and Images (SIBGRAPI), 2017.