As humans, we encounter countless objects daily. We effortlessly recognize object across variations despite the fact that the objects might vary in size, scale, translation or rotation. Humans can identify previously seen objects and posses the ability to learn new instances with minimal or no supervision. Human visual system continues to learn and adapt to ever changing surroundings. In recent years, there have been significant advances in the field of computer based recognition systems. While significant strides have been made towards building automated recognition systems, these systems face multiple challenges when operating in evolving environments. Operational issues such as changing data distributions, perturbations in input/output conditions and ever changing requirements of the system users, pose challenges in operational environments. In this work we highlight specific operational challenges such as handling partial information, incremental model adaptation, large-scale classification and propose solutions towards addressing these challenges iii Dedication This thesis is dedicated to my family for their love, constant support, encouragement and patience. iv Acknowledgements

Increasing access to large, non-stationary face datasets and corresponding demands to process, analyze and learn from this data. This has lead to a new class of on-line/incremental face recognition problems. While it is ad-vantageous to build large scale learning systems when re-sources permit, a counter problem of learning with limited resources in presence of streaming data arises. We present a budgeted incremental support vector learning method suit-able for online learning applications. Our system can pro-cess one sample at a time and is suitable when dealing with large streams of data. We discuss multiple budget mainte-nance strategies and investigate the problem of incremen-tal unlearning. We propose a novel posterior probability estimation model based on Extreme Value Theory (EVT) and show its suitability for budgeted online learning ap-plications (calibration with limited data). We perform thor-ough analysis of various probability calibration techniques with the help of methods inspired from meteorology. We test our methods on Labeled Faces in the Wild dataset and show suitability of the proposed approach for face verifica-tion/recognition 1

In this paper, we propose a method to apply the popular cascade classifier into face recognition to improve the com-putational efficiency while keeping high recognition rate. In large scale face recognition systems, because the proba-bility of feature templates coming from different subjects is very high, most of the matching pairs will be rejected by the early stages of the cascade. Therefore, the cascade can im-prove the matching speed significantly. On the other hand, using the nested structure of the cascade, we could drop some stages at the end of feature to reduce the memory and bandwidth usage in some resources intensive system while not sacrificing the performance too much. The cascade is learned by two steps. Firstly, some kind of prepared features are grouped into several nested stages. And then, the thresh-old of each stage is learned to achieve user defined verifi-cation rate (VR). In the paper, we take a landmark based Gabor+LDA face recognition system as baseline to illus-trate the process and advantages of the proposed method. However, the use of this method is very generic and not lim-ited in face recognition, which can be easily generalized to other biometrics as a post-processing module. Experi-ments on the FERET database show the good performance of our baseline and an experiment on a self-collected large scale database illustrates that the cascade can improve the matching speed significantly. 1.