Person re-identification to bridge domain gaps

Abstract

With a dramatic increase in CCTV cameras, person re-identification (Re-ID), which is the most promising application of intelligent video surveillance, has been studied widely. Its objective is to identify a specific person across non-overlapping cameras under various locations, and Re-ID plays a pivotal role in providing security to the public due to its usefulness and practical importance. Although some Re-ID methods have shown impressive performance in the limited environments, all of the domain issues in Re-ID systems have not been explored sufficiently. In this dissertation, we study various domain issues that may occur in real video surveillance situations and focus on effective person re-identification techniques to bridge domain gaps. The first part is a visible-infrared person re-identification (VI-ReID) task, which aims to match pedestrians observed from visible and infrared cameras with different spectra. Since visible light cameras can not capture all the appearance characteristics of a person under poor illumination conditions, most surveillance cameras automatically switch from visible to infrared mode in dark environments. Thus, it becomes essential to consider the VI-ReID task. For this task, it is vital to reduce the domain gap between visible and infrared images. We propose a Hierarchical Cross-Modality Disentanglement (Hi-CMD) method, which automatically disentangles ID-discriminative factors and ID-excluded factors from visible-thermal images. We only use ID-discriminative factors for robust cross-modality matching without ID-excluded factors such as pose or illumination. Extensive experimental results demonstrate that our method outperforms the state-of-the-art methods on two VI-ReID datasets. The second part is a domain generalizable person re-identification task, which aims at learning a generalizable Re-ID model only using a given source domain so that the trained model can be applied robustly in any target domain. This task is much more plausible for real-world applications since the trained model can be utilized in any situation without updates. The core issue is to bridge a seen source domain and an unseen target domain. To this end, we propose a Meta Batch-Instance Normalization (MetaBIN) method. Our main idea is to generalize normalization layers by simulating unsuccessful generalization scenarios beforehand in the meta-learning pipeline. Our MetaBIN framework prevents our model from overfitting to the given source styles and improves the generalization capability to unseen domains. Extensive experimental results show that our method outperforms the state-of-the-art methods on the large-scale domain generalization Re-ID benchmark and the cross-domain Re-ID problem.