Camera-aware Progressive Learning for Unsupervised Person Re-Identification.

Yuxuan Liu,Hongwei Ge,Liang Sun,Yaqing Hou

DOI: https://doi.org/10.1007/s00521-023-08301-w

2023-01-01

Neural Computing and Applications

Abstract:Unsupervised person re-identification (Re-ID) is a challenging task due to the problems of the view-specific gap across different cameras and the missing of ground truth labels. The intra-class variance caused by the camera domain gap makes the model difficult to distinguish cross-view discriminative information. The noisy pseudo labels may be misguide the model toward deteriorated over-fitting. Improving the clustering accuracy of pseudo labels and alleviating the intra-class variance can improve the performance of unsupervised person Re-ID. Most methods improve clustering accuracy by optimizing the clustering distribution of instance features in the memory bank. However, the optimized distribution is still unbalanced and has a large intra-class gap. In order to address this problem, we propose a camera-aware progressive learning (CAPL) for unsupervised person re-identification. In CAPL, on the one hand, clustering edge features are explored and used to update all instance features belonging to the same pseudo label in the memory bank, which makes the instance feature distribution more balanced; on the other hand, camera-relevant features are explored to update all instance features with the same pseudo-labels in the memory bank as well as the corresponding camera class centroids, which makes the model effectively reduces the intra-class variance. This process is performed iteratively, and finally, the global clustering distribution can be achieved progressively. Extensive experiments on four person Re-ID benchmarks and a vehicle Re-ID benchmark demonstrate that our proposed approach outperforms the state-of-the-art methods in terms of mAP and CMC.