Unsupervised Domain Adaptive Learning via Synthetic Data for Person Re-identification

Noname manuscript No. (will be inserted by the editor) Abstract Person re-identification (re-ID) has gained more 1 Introduction and more attention due to its widespread applications in intelligent video surveillance. Unfortunately, the mainstream re-identification(Re-ID) aims at identifying images of the deep learning methods still need a large quantity of labeled same pedestrian across non-overlapping camera views in different data to train models, and annotating data is an expensive places, which has attracted a lot of research interests work in real-world scenarios. In addition, due to domain since the urgent demand for public safety and the increasing gaps between different datasets, the performance is dramatically number of surveillance cameras. Benefiting from the development decreased when re-ID models pre-trained on label-rich of deep learning (He et al., 2016; Szegedy et al., datasets (source domain) are directly applied to other unlabeled 2015) and the availability of labeled re-ID datasets Zheng datasets (target domain). In this paper, we attempt to et al. (2015), (Ristani et al., 2016; Wei et al., 2018; Ye et al., remedy these problems from two aspects, namely data and 2020), CNN-based re-ID methods (Ye et al., 2020; Yin et al., methodology. Firstly, we develop a data collector to automatically 2020; Zhu et al., 2021; Tao et al., 2013; Li et al., 2017), generate synthetic re-ID samples in a computer have made remarkable performance improvements in a supervised game, and construct a data labeler to simultaneously annotate manner. However, the aforementioned approaches them, which free humans from heavy data collections need a multitude of accurately labeled and diversified data to and annotations. Based on them, we build two synthetic person learn the discriminative features, and the current datasets are re-ID datasets with different scales, "GSPR" and "mini-not able to perfectly satisfy the demands in dataset scale or GSPR" datasets. Secondly, we propose a synthesis-based data diversity.