判别性信息增强的行人再识别

Person Re-identification by Discriminative Information Enhancement

在行人再识别中,由相机风格及视角变化造成的行人外观差异严重影响了模型的性能.为缓解该现象,提出一种判别性信息增强的行人再识别方法.该方法由辅助子网络、主干子网络以及时空信息嵌入三部分构成.首先,在辅助子网络中实现不同视角行人图像的风格转换以减少风格差异对性能的影响.为增强特征的判别性,将原图特征和迁移图像特征进行拼接.此外,在主干子网络中通过局部特征迫使主干子网络在关注全局特征的同时,能更多利用局部具有鉴别性的信息.最后,引入行人的时空信息来缓解难样本对识别性能造成的影响.通过实验证明所提算法性能优于大部分主流方法,消融实验也验证了所提算法的有效性.

In person re-identification,the difference in appearance of person caused by camera style and viewpoints changes seriously effects the performance of the model.To alleviate this phenomenon,a person re-identification method based on discriminative information enhancement is proposed.The method consists of three parts:auxiliary sub-network,backbone sub-network and spatial-temporal information embedding.First,realize the style conversion of person images from different perspectives in the auxiliary sub-network to reduce the impact of style differences on performance.In order to enhance the discriminability of features,the original image features and the transfer image features are concatenated together.In addition,local features are used in the backbone sub-network to force the backbone sub-network to pay more attention to global features while making more use of local discriminative information.Finally,the person spatial-temporal information is introduced to alleviate the impact of difficult samples on recognition performance.Experiments prove that the performance of the proposed algorithm surpassing the state of the art,and ablation study also verify the effectiveness of the proposed algorithm.