Person Re-Identification (ReID) aims at retrieving the images of a query person from a large set of gallery images. It has been an attractive research field in computer vision due to the ever-increasing demand for camera networks in public spaces. In recent years, significant improvements have been observed in person ReID task in parallel with the developments in deep learning. However, due to the large discrepancy between the training/test distributions, the ReID models generally lack in generalizing to the test data, which is the phenomenon known as overfitting.
In this thesis, we propose an ensemble method to increase the generalization capability of the ReID models. Ensemble models, which consist of multiple base learners whose decisions are combined in test time, deal with the overfitting problem effectively and increase the generalization capability. However, training an ensemble of deep networks is computationally inefficient. To overcome this difficulty, we create diverse and accurate base learners in a single network by designing a multi-branch architecture. Detailed analysis of the experiments on three benchmark datasets demonstrates the effectiveness of our approach, which outperforms the state-of-the-art approaches. We adapt the proposed approach to Binary Neural Networks. Our experiments show that the proposed approach improves the Binary Neural Networks in terms accuracy and training stability in image classification task and outperforms the conventional ensemble model by a large margin in person ReID, which indicates that our model is not only an ensemble model, but also an effective regularizer for deep networks. |