Reliability Exploration with Self-Ensemble Learning for Domain Adaptive Person Re-identification

Person re-identifcation (Re-ID) based on unsupervised domain adaptation (UDA) aims to transfer the pre-trained model from one labeled source an unlabeled target domain. Existing methods tackle this problem by using clustering generate pseudo labels. However, labels produced these techniques may be unstable and noisy, substantially deteriorating models’ performance. In paper, we propose a Reliability Exploration with Self-ensemble Learning (RESL) framework for adaptive person ReID. First, increase feature diversity, multiple branches are presented extract features different data augmentations. Taking temporally average as mean teacher model, online label refning is conducted its dynamic ensemble predictions soft Second, combat adverse effects of unreliable samples in clusters, sample reliability estimated evaluating consistency clusters’ results, followed selecting reliable instances training re-weighting contribution within Re-ID losses. A contrastive loss also utilized cluster-level memory which updated feature. The experiments demonstrate that our method can signifcantly surpass state-of-the-art performance