Science Deep learning for person re - identification

Person re-identification is the task of ranking a gallery of automatically detected images of persons using a set of query images. This is challenging due to the different poses, viewpoints, occlusions, camera configurations, image distortions, lighting conditions, image resolutions and imperfect detections, which all affects a person re-identification system’s performance. Recently deeply learned systems have become prevalent in the person re-identification field as they are capable to deal with the various obstacles encountered. One such a system is ConvNet using a coarse-to-fine search framework (ConvNet+C2F), which is developed with both a high retrieval accuracy as a fast query time in mind. In this thesis we propose several adaptations to ConvNet+C2F to improve its performance. We use the novel convolutional model Xception to construct a new ConvNet called XConvNet, train it using the modern Adadelta model optimizer and demonstrate that a smaller coarse descriptor improves retrieval time and accuracy for C2F. With the proposed improvements XConvNet+C2F achieves state-of-the-art results on two different well known datasets for person re-identification, i.e. Market-1501 and CUHK03. Furthermore, we investigated the possibilities of an artificial extension of the training set using generated images and study the effect of different batch sizes used in the classifier training.