Face recognition in unconstrained images and videos
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
... le 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. ATTENTION: The Singapore Copyright Act applies to the use of this document. Nanyang Technological University Library ii Acknowledgments Firstly, I would like express my utmost appreciation to my supervisor, Assoc Prof Xu Dong, for his time, support and guidance throughout the entire master programme. The journey with him has been an enjoyable and fruitful learning experiment; his ideas and guidance always keep my research in the right direction. Without his patience, I could not have completed the work in this thesis. I would like to thank all my research project members and all colleagues in CeMNet centre, especially Mr.