In-depth exploration of attribute information for person re-identification
- PDF / 3,734,909 Bytes
- 16 Pages / 595.224 x 790.955 pts Page_size
- 39 Downloads / 221 Views
In-depth exploration of attribute information for person re-identification Jianyuan Yin1 · Zheyi Fan1 · Shuni Chen1 · Yilin Wang1
© Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Pedestrian’s attribute information plays an important role in person re-identification (re-ID) for its complementary to pedestrian’s identity labels. However, there are few methods to utilize attribute information, which limits the development of re-ID community. In this paper, we analyze the effect of attribute information on re-ID to obtain both qualitative and quantitative results, indicating the potential for in-depth exploration of attribute information. On this basis, we propose an Identity Recognition Network (IRN) and an Attribute Recognition Network (ARN). IRN enhances the attention to pedestrian’s local information while identifying pedestrians’ identity. ARN calculates the attribute similarity among pedestrians accurately to promote the identification of IRN. The combination of them makes deep exploration of attribute information and is easy to implement. The experimental results on two large-scale re-ID benchmarks demonstrate the effectiveness of our method, which is on par with the state-of-the-art. In the DukeMTMC-reID dataset, mAP (rank-1) accuracy is improved from 58.4 (78.3) % to 66.4 (82.7) % for ResNet-50. In the Market1501 dataset, mAP (rank-1) accuracy is improved from 75.8 (90.5) % to 79.5 (92.8) % for ResNet-50. Keywords Person re-identification · Attribute · Identity recognition network · Attribute recognition network
1 Introduction Re-ID [54] is a cross-camera retrieval task. Given a query person-of-interest, it aims at retrieving images containing the same person from a gallery collected by multiple cameras. It is a meaningful problem not only because its significant applications in real world but also for its integration with other computer vision tasks, such as object tracking [6, 7]. In spite of the impressive achievements by the re-ID community, the performance of re-ID still has a large margin for improvement. Most existing methods treat re-ID as a multiclassification task: we can train the identification model
All authors contributed to the study conception and design. All authors read and approved the final manuscript. Zheyi Fan
[email protected] 1
School of Information and Electronics, Institute of Signal and Image Processing, Beijing Institute of Technology, 5 South Zhongguancun Street, Haidian District, Beijing, 100081, China
through the identity labels of pedestrians. However, the identity labels cannot take into account the local information of pedestrians, which will confuse the identification model. As show in Fig. 1, when we only identify pedestrians from the identity dimension, although the model can distinguish between pedestrians with large gaps, it is hard to succeed again if two pedestrians are similar in the dimension of identity. For example, it is difficult to distinguish id5 and id6 in the dimension of identity, although they are different in genders. Ins
Data Loading...
