Set-to-Set Face Recognition Under Variations in Pose and Illumination

We present a face recognition method using multiple images where pose and illumination are uncontrolled. The set-to-set framework can be utilized whenever multiple images are available for both gallery and probe subjects. We can then transform the set-to-set classification problem as a geometric one by realizing the linear span of the images in a given resolution as a point on the Grassmann manifold where various metrics can be used to quantify the closeness of the identities. Contrary to a common practice, we will not normalize for variations in pose and illumination, hence showing the effectiveness of the set-to-set method when the classification is done on the Grassmann manifold. This algorithm exploits the geometry of the data set such that no training phase is required and may be executed in parallel across large data sets. We present empirical results of this algorithm on the CMU-PIE database and the extended Yale face database B, each consisting of 67 and 28 subjects, respectively.