基于域鉴别网络和域自适应的行人重识别

Person re-identification based on domain discriminative network and domain adaptation

针对行人重识别无监督跨域迁移问题,提出一种基于域鉴别网络和域自适应的行人重识别算法。首先,使用改进ResNet-50训练监督域鉴别网络模型,加入共享空间组件得到特征不变属性,用于区分类间图像,并基于对比损失和差异损失来提高模型的分类性能。其次,利用域自适应无监督迁移方法由源域数据集导出特征不变属性,并应用到未标记的目标域数据集上。最后,匹配查询图像和共享空间中的图库图像执行跨域行人重识别。为验证算法有效性,在CUHK03、Market-1501和DukeMTMC-reID数据集上进行了实验,算法在Rank-1准确度分别达到34.1%、38.1%和28.3%,在mAP分别达到34.2%、17. 1%和17.5%,最后还验证了模型各个组件在训练阶段的必要性。结果表明本文算法在大规模数据集上的性能优于现有的一些无监督行人重识别方法,甚至接近于某些传统监督学习方法的性能。

Aiming at the problem of unsupervised cross-domain migration for person re-identification,an algorithm based on domain discriminative network and doma in adaptation is proposed.Firstly,the improved ResNet-50training supervised domain discrimina tive network model is used,and the shared space component is added to obtain the feature invariant attribute,which is used for the inter-class classification image,and the classification perfor mance of the model is improved based on the contrast loss and the difference loss.Secondly,the do main invariant migration method is used to derive the feature invariant attribute from the sour ce domain dataset and should be on the unmarked target domain dataset.Finally,the matching query image and the gallery image in the shared space perform cross-domain person re-identificatio n.In order to verify the validity of the algorithm,experiments were carried out on the CUHK03,Market-1501and DukeMTMC-reID datasets.The accuracy of the algorithm in Rank-1reached 34.1%,38.1% and 28.3%, respectively,and reached 34.2%,17.1% and 17.5% in mAP.Finally,the necessity of each component of the model in the training phase is verified.The results show that the perfor mance of the proposed algorithm on large-scale datasets is better than some existing unsuper vised person re-identification methods,even close to The performance of some tr aditional supervised learning methods.