Navigation auf uzh.ch
Navigation auf uzh.ch
Automatic face recognition has been a hot research topic and in the recent years, face recognition algorithms have matured into a stage where they are used in many daily activities. For example, people unlock their phones using face recognition, or airports rely on automatic face recognition to enter through e-gates. All of these applications, however, require to compare frontal faces and fail when a half-profile or even profile face is shown.
In earlier times, face recognition across pose has been achieved through rank lists such as in Learning Invariant Face Recognition from Examplesor Pose, Illumination and Expression Invariant Pairwise Face-Similarity Measure via Doppelgänger List Comparison. The main assumption here is that two profile faces look similar if and only if they look similar in frontal pose, too. Based on a model database containing examples of frontal and non-frontal faces, we represent a face image as a list of similarities to the faces in the model database that have the same pose. Finally, we transform this similarity list into a rank list, i.e., the most similar person gets rank one, the second most rank two and so on. In order to compare two faces with each other, their respective rank lists are compared using specifically dedicated rank list similarity functions.
Originally, the idea is to only compare a frontal face to all frontal faces in the model database, and a profile face only to other profile faces in the model database. However, this requires to know the pose of the face in the images, which might be failure-prone. The research in this Bachelor's thesis focuses on the extension of the rank list comparison to use more modern face recognition approaches, i.e., deep learning-based face recognition. Using a pre-trained network, for example FaceNet, we extract deep features from faces. The deep features of some of the subjects are put into our model gallery, while other peoples' faces are used for testing. Instead of comparing only frontal to frontal and profile to profile faces, we compare the incoming face to all faces in the model database and take the maximum similarity per person to compute our initial similarity list that we later turn into a rank list. We follow the evaluation protocol defined in Cross-pose LFW: A Database for Studying Cross-Pose Face Recognition in Unconstrained Environments which allows us to directly compare our results to the state of the art.
