Noise-robust re-identification with triple-consistency perception

Xueping Wang, Zhanpeng Shao, Shixi Luo, Jiazheng Wang, Min Liu, Jianhua Dai, Jun Cheng

Abstract: Highlights•A self-consistency strategy is proposed to refine our model and avoid it overfitting to the noisy labels at the beginning of model training by mining the consistency of annotations and predictions.•A context-consistency loss is presented to enhance the robustness of the proposed model by exploring the context relations and encouraging each sample to have similar predictions to its neighbors.•A cross-view consistency loss is introduced to encourage the consistency of samples across different views by minimizing the Jensen-Shannon divergence of the predictions from different views.•Experimental results validate the superiority of the proposed TcP-ReID model over the competing methods, on Market1501, our method achieves 85.8% in rank-1 and 56.3% in mAP under instance-independent noise with noise ratio 50%.