TGC20ReId: a dataset for sport event re-identification in the wild

Abstract

Person re-identification (Re-ID) is the task of retrieving a person of interest taken from different cameras or from the same camera in different occasions. To address this challenging task, a large amount of labelled data is required both for testing and for learning. Such high quality annotated data is still rare for many Re-ID applications. In this paper, we introduce a novel dataset to evaluate Re-ID methods in complex real-world scenarios. In this case, we will be using a sporting event as the scenario. For this aim, participants in a 128 km night and day course were captured in five different recording points along the track and later manually identified and annotated. The dataset is evaluated using state-of-the-art techniques and wide array of experimental setups are considered, such as: day vs. night, motion blur, changes in clothing, etc. The evaluation suggests that the features present in the proposed dataset (time difference, unrestricted weather and illumination capture conditions, and the possibility of clothes changes between probe and gallery) pose a challenging scenario for future development in the Re-ID problem.