Person re-identification is the task of searching for images of the same person taken with different cameras. Since it has the potential to be introduced into a variety of applications, such as tracking people in smart retail and finding lost children, the field of person re-identification has attracted attention not only in academia but also in industry. To date, research has been conducted on the application of deep learning to person re-identification, achieving accuracy levels that far exceed methods that use manually generated features. However, models trained in a labeled learning environment (source domain) have the problem of significantly degrading accuracy when applied to a different environment (target domain). As a method to solve this degradation in accuracy, unsupervised domain adaptation methods that aim to transfer knowledge from the source domain to the target domain have attracted attention.

Many conventional methods take an approach of assigning pseudo-labels to unlabeled data through clustering, but the key is how to identify images of the same person that are difficult to recognize as the target person (Hard-Positive), while eliminating images of other people who look like the target person (Hard-Negative). Many conventional methods utilize a convolutional network structure that uses only a single output generated from Global Average Pooling (GAP) to extract global features of the input and have the problem of not taking into sufficient consideration local features that are effective in fine-grained tasks.

In this study, the authors proposed a new learning method that uses the output of Global Max Pooling (GMP) to extract local features in addition to GAP, which extracts global features. Additionally, the proposed learning method considers the difference between GAP and GMP output, which have different output characteristics.

In the proposed method, the author experimentally demonstrated that the discrimination of Hard-Negatives is improved while maintaining the recognition accuracy of Hard-Positives by using the intersection set of pseudo label sets obtained by clustering the outputs of GAP and GMP individually. The effectiveness of the proposed method is also demonstrated by the fact that its accuracy exceeds that of representative conventional methods by conducting domain adaptation experiments using large-scale datasets, Market-1501 and MSMT17. The human cost of annotation for each implementation environment, which is a problem when putting person re-identification into practical use, can be solved by using the proposed method, which shows the great contribution of this paper.