Noise Resistible Network for Unsupervised Domain Adaptation on Person Re-Identification

Unsupervised domain adaptation on person re-identification (re-ID), which adapts the model trained source dataset to target dataset, has drawn increasing attention over past few years. It is more practical than traditional supervised methods when applied in real-world scenarios since they require a huge number of manual annotations specific domain, unrealistic and even under personal privacy concerns. Currently, pseudo label-based method one most promising solutions this area. However, such methods, label noise ignored remains challenge hindering further performance improvements. To solve problem, paper proposes novel unsupervised re-ID framework named Noise Resistible Network (NRNet), mainly consists two dual-stream networks. For thing, during generation, NRNet utilizes network, denoted as clustering generate discriminative features unseen for clustering, reducing noise. another, avoid problem close loop amplification conventional other network temporally average constructed outside learn how identify images same person. In addition, networks are designed with guiding mechanism, allows shallow representative feature embedding from deep network. Extensive experimental results widely-used benchmark datasets, i.e., Market-1501 DukeMTMC-reID demonstrate that our proposed outperforms state-of-the-art methods.