Unsupervised domain adaptation is a challenging task in person re-identification (re-ID). Recently, cluster-based methods achieve good performance; clustering and training are two important phases in these methods. For clustering, one major issue of existing methods is that they do not fully exploit the information in outliers by either discarding outliers in clusters or simply merging outliers. For training, existing methods only use source features for pretraining and target features for fine-tuning and do not make full use of all valuable information in source datasets and target datasets. To solve these problems, we propose a Threshold-based Hierarchical clustering method with Contrastive loss (THC). There are two features of THC: (1) it regards outliers as single-sample clusters to participate in training. It well preserves the information in outliers without setting cluster number and combines advantages of existing clustering methods; (2) it uses contrastive loss to make full use of all valuable information, including source-class centroids, target-cluster centroids and single-sample clusters, thus achieving better performance. We conduct extensive experiments on Market-1501, DukeMTMC-reID and MSMT17. Results show our method achieves state of the art.
Unsupervised domain adaptation in person re-identification is a challenging task due to domain shifts between different datasets. Some cluster-based domain adaptation methods have made good progresses. However, most of these methods only focus on differences between global features while ignoring local features. In this paper, we propose a Patch and Cluster-based Domain adaptation (PCD) method. Specifically, we introduce the concept of patch-level local features and use spatial transformer to generate patches. We consider both global features (image-level) and local features (patch-level), so we get great accuracy. We evaluate our model on Market-1501 and DukeMTMC-reID. Extensive experiments show the efficiency of the method.
3 March 2022 in Journal of Circuits, Systems and Computers