Resumen de Object re-identification in urban scenarios
Paula Moral de Eusebio
Object re-identifcation plays a pivotal role in the development of automated image-based applications by facilitating the association of images depicting the same object identity across diferent cameras or time instants. The main challenges present in this task are related to large variations in lighting, diverse poses, diferent viewpoint, appearance and occlusions. One of the most prominent applications of re-identifcation focuses on enhancing public safety through person re-identifcation in video surveillance. However, this thesis is focused on urban element re-identifcation, aiming to propose technologies that are applicable to monitoring diferent urban objects. The idea of re-identifying people was not followed in order to avoid privacy conficts with the use of personal data. The initial phase involves leveraging an existing dataset to explore vehicle re-identifcation. As the diversity of datasets containing urban objects is very limited, the research extends to study the entire re-identifcation system, starting with dataset creation. The thesis begins with a comprehensive review of state-of-the-art re-identifcation techniques and datasets. Detailed descriptions are provided for the methodologies employed, including feature extraction techniques utilizing neural network architectures and ensemble methods that integrate image-and video-based features. Furthermore, the research employs diverse post-processing strategies to refne re-identifcation accuracy. Subsequently, we propose to enhance performance through the incorporation of synthetic data and the application of unsupervised domain adaptation techniques. These approaches aim to reduce the gap between test and training data by generating pseudo-labels in test set. Additionally, we provide an evaluation framework because the vehicle dataset used does not provide the ground truth of the test set and the evaluation server used during the competition has not always been available after its completion. One of the overall objectives of this thesis is to comprehensively examine an end-to-end re-identifcation system. We have focused on urban elements, starting with data acquisition processes. Additionally, the research involves the exploration of various object detection algorithms to obtain precise bounding boxes for each object identity. Subsequently, these bounding boxes are utilized as inputs for evaluating the performance of re-identifcation systems of urban elements
