Person Re-identification Robustness Research on XQDA

Abstract

Person re-identification (re-id) has become increasingly popular in the community due to the importance of surveillance and security. Some re-id benchmark datasets have been provided to evaluate re-id algorithms, such as VIPeR, CUHK01, CUHK03. These datasets are often labeled by human operators and there are no error labeling under ideal conditions. However, when we apply re-id algorithm in practice, the detector must be used and the detected pedestrians are usually labeled by machine automatically to augment training datasets for promoting re-id performance. In this situation, error labels may be brought into the dataset. The VIPeR dataset is a small scale that has been evaluated mostly. When we use it as the basic training dataset, error labels will make a big effect on the re-id precision. As an important hand-crafted algorithm, XQDA (Cross-view Quadratic Discriminant Analysis) focuses on handling the multiple cameras for the same identity under the different point of views. It can achieved a good performance with fast computing speed and high precision on the datasets mentioned above. In this paper, we focus on studying robustness of XQDA under different levels of noise on VIPeR. Then we try to propose some methods to relieve the bad effect due to the error labels, enhancing the robustness further.