In recent years, face recognition remains as one of the most attractive research topic in the computer vision field. Many methods have been proposed to deal with the large appearance changes of human face, yet it is still debatable which method works best on the unconstrained environment in ordinary photograph and video. This thesis thoroughly reviews and implements some state-of-the-art methods for image-based face recognition as well as completely studies their performance on various large scale dataset. Motivated from the results on image datasets, a novel framework is proposed for face verification in unconstrained videos. By leveraging the pose angle information, we employ a divide and conquer approach through the following steps (a) divide the keyframes of the original video into several pose categories from extreme left to right profile face and synthesize the appearance at all missing poses (b) propose the ensemble cross-pose classifiers to recognize human faces despite pose differences by cross reference the subset of the original training data with the same pose categories. Extensive experiments on the large-scale YouTube video dataset clearly demonstrate the effectiveness and robustness to pose variation of our proposed framework.