An unsupervised domain adaptation concept for face recognition applications

When facing real world data, domain shift is a significant challenge for face recognition methods. A domain mismatch between the datasets used to train a recognition model in the lab and the target data leads to a negative impact on the performance of the recognitions system. Because it is difficult for some domains, like surveillance footage, to collect sufficiently large training datasets, a concept to generalize from data originating from a set of given domains to novel domains is proposed. The critical difference to most existing learning strategies under these conditions is the additional constraint that no data is available from the target domain which is called unsupervised domain adaptation. The suggested concept combines large-margin dimension reduction with an SVM-based framework for unsupervised domain adaptation. This unified strategy benefits by using more of the available knowledge about the training data in the sense of dataset/domain membership of the training samples compared to simply combining all available training datasets into one large dataset.