Navigation auf uzh.ch
Navigation auf uzh.ch
One use case of automatic face recognition is user authentication, e.g., someone tries to unlock their phones by comparing the live camera image to a previously recorded template that is stored on the phone. Facial attributes have the possibility to act as soft biometrics, i.e., they enable to reject an access attempt if it detects that, e.g., the gender of the stored template and the live face differ. While face recognition usually requires high-resolution imagery to work appropriately, most facial attributes can be extracted from low-resolution and badly illuminated images.
In MOON: A Mixed Objective Optimization Network for the Recognition of Facial Attributes, we showed that face recognition can be achieved by extracting 40 different facial attributes from a face, and using this vector of attributes for face recognition. However, there are many attributes that are not or only vaguely related to identity, such as the blurriness of an image, hair color, wearing hats or makeup and alike. While in MOON, we disregarded this fact and simply used all 40 attributes in exactly the same way, it is to be expected that an automatic analysis of the stability of such attributes -- and also the stability of the automatic prediction of such attributes -- will enable a significant improvement of the authentication success. Additionally, it would be important to check how the stability is influenced based on the presence or the absence of an attribute: While preliminary experiments showed that, for example, the absence of the attractive flag can be very stable for a given person, the presence of this flag for another person can vary drastically based on factors such as illumination, facial expression, make-up, or alike. The design of a similarity function should incorporate such effects.
Experiments will be executed on the CelebA dataset to design an attribute similarity function that is better suited for face recognition, and on the Labeled Faces in the Wild (LFW) dataset to compare results to the ones previously published in the MOON paper.
