融合交互性特征信息的余弦度量行人重识别

Person re-identification with cosine metric fusing interactive feature information

针对行人重识别中难以表达特征间相关性信息,以及欧氏距离度量的损失函数忽略了特征向量角度影响的问题,提出一种融合交互性特征的余弦距离度量行人重识别网络。通过聚合分组注意力模块提取的特征,实现不同子特征跨通道的整合交互。训练阶段采用余弦度量的三元组损失,结合特征空间的批量规一化操作,消除模长波动的影响,从角度维度判别行人差异。采用广义平均池化,保留更完整的特征信息。在Market-1501和DukeMTMC-reID两个数据集的实验结果表明,网络能有效提升行人重识别精度。

To solve the problems that it is difficult to express the correlation information between features in person re-identification,and the loss function of Euclidean distance metric ignores the problem of pedestrian embedded feature vector angle,a cosine distance metric person re-identification network integrating interactive features was proposed.By aggregating the features extracted by the grouped attention module,the integrated interaction of different sub-features across channels was realized.In the training phase,the triplet loss of cosine metric was used,combined with the batch normalization operation of the feature space,to eliminate the influence of modulo-length fluctuations,and to discriminate pedestrian differences from the angle dimension.Generalized-mean pooling was also employed to preserve more complete feature information.Experiments on the Market-1501 and DukeMTMC-reID datasets show that the network can effectively improve the accuracy of person re-identification.