Proceedings of The 7th International Conference on Knowledge and Innovation in Engineering, Science and Technology
Year: 2020
DOI:
Refined Granularity Extraction for Person Reidentification
Motaz Sabri
ABSTRACT:
The usage of global and partial features has proven essential in person reidentification (Re-ID) tasks. Extracting both features does not happen uniformly resulting in representations that are either focusing on local representations or posture details. This consequently lowers the efficiency and robustness against scenarios with large variances. In this paper, we propose a feature learning strategy integrating discriminative information with refined granularities. We carefully design a multibranch deep network architecture with one expert branch for global feature representations and two expert branches for local feature representations. We replace focusing on semantic regions with a focus on several stripes of images partitions. The stripes number varies in different local branches to obtain local feature representations with multiple granularities. We evaluate our approach on four challenging datasets (Market1501, MSMT17, DukeMTMC Re-ID and CUHK03) where it achieves state of the art performance among both supervised and unsupervised methods.
Keywords: Discriminative features learning, Feature localization, Granularity learning, Metric learning, Person Reidentification.