Face Recognition in Adverse Conditions: A Look at Achieved Advancements

The Role Of Pose

Pose variation is one of the great challenges for robust face recognition. We can divide techniques to address this problem in two broad categories: those only relying on 2D images and those using 3D models.

Methods based on 2D images were the first to be investigated (see for example Reisfeld and Yeshurun, 1998). In their simplest basic versions, those methods perform face normalization based on the location of the eyes and mouth. A two dimensional affine transformation is uniquely determined by three points. Warp is defined as the affine mapping of a face image determined by the location of the eyes and mouth in the given image and by the usual standard locations. In a more general case, if two given samples are well aligned, there exists an approximated linear mapping between the two images of the person captured under variable poses. In order for this mapping to be consistent for all subjects, it is required that their facial images are aligned pixel-wise. Since this kind of alignment is a problem in itself, theone actually performed usually relies on very few facial landmarks, such as the two eye centers. However, in this way face images are aligned quite coarsely, so that the assumption of linear mapping no longer holds theoretically.