Relevance Transfer: Towards Robust Distillation in Person Re-Identification

Jing Zhou,Zuliu Yang,Junting Yang,Yong Zhao
DOI: https://doi.org/10.1088/1742-6596/1914/1/012011
2021-01-01
Journal of Physics: Conference Series
Abstract:Abstract Person Re-Identification(ReID) is developing rapidly these years powered by deep learning. However most existing methods with state-of-the-art performance are highly dependent on “heavy” CNN backbones, which are hard to deploy on lightweight computing devices. In this paper, we propose a novel feature distillation method called Relevance Transfer, which aims to transfer the knowledge hidden inside the relation of different feature parts from teacher network to tiny student network. And we conduct a Multi-Granularity Horizontal Partition (MGHP) strategy to make training more efficient and robust. As results on two mainstream ReID datasets show, our method achieves competitive performance with original logits Knowledge Distillation (logits KD), and gains even more aggressive performance when works along with logits KD.
What problem does this paper attempt to address?