Pose unconstrained face recognition based on SIFT and alignment error

Pose variation is one of the key challenges for practical face recognition problem. Face recognition under well-controlled settings, like frontal face and good illumination, achieved high performance. But it fails when they are directly adopted to face recognition with large pose change. In this paper, we propose a novel framework using the combination of SIFT and alignment error (SIFT-AE) to perform pose invariant face recognition. SIFT (Scale Invariant Feature Transformation) is an effective local descriptor for face recognition under small pose change, which is scale and rotation invariant as well. However, the performance declined in case of large pose variance. To compensate this declination, Lucas-Kanade method is used to align the probe image and the gallery image, and the alignment error is deducted from the number of matching for SIFT algorithm. This alignment error provides additional information even in case of large pose change, while it is not distinctive alone. Therefore, the combination of SIFT and alignment error gains the performance for face recognition with large pose variance. Experiment results show our algorithm achieves impressive improvement compared with either SIFT or online alignment.