基于体形引导的无监督换装行人重识别算法

Shape-guided Cloth-changing Unsupervised Person Re-identification Algorithm

在无监督换装行人重识别任务中,服装的变化会造成外观特征表示不稳定,导致网络的识别率降低。为此,基于协同对齐互交叉注意力(Co-attention Aligned Mutual Cross-attention, CAMC)算法提出了一种由行人体形引导的双分支网络(Shape-guided Network, SG-Net)。SG-Net提取行人的外观特征和关键点信息,通过注意力机制增强外观特征中的脸部区域,将关键点信息编码为人体的体形特征;外观特征与体形特征对齐后,将与服装无关的身份信息传递到外观特征空间中,生成鲁棒的外观特征表示;无监督场景下,融合外观特征与体形特征进行聚类为样本生成伪标签用于监督网络训练。在Celeb-ReID-light数据集实验结果表明,与基准网络ResNet50相比,SG-Net算法的mAP和Rank-1指标分别提高了2.3%和5.0%。

In the task of unsupervised cloth-changing person re-identification,the change of clothing caused the appearance feature representation to be unstable,which led to the reduction of the recognition rate of the network.Therefore,based on the Co-attention Aligned Mutual Cross-attention(CAMC)algorithm,a Shape-guided double-branch Network(SG-Net)was proposed.SG-Net extracts the appearance features and keypoint information of pedestrians,enhances the face area in appearance features through the attention mechanism,and encodes the keypoint information into the shape information of human body.After the appearance feature aligned with the shape feature,the identity information unrelated to clothing was transferred to the appearance feature space to generate a robust appearance feature representation.In the unsupervised scenes,the appearance features and body shape features were integrated to cluster samples and generate pseudo-labels,which were used for supervising network training.The experimental results on the Celeb-ReID-light dataset show that compared with the benchmark network ResNet50,the mAP and Rank-1 indexes of SG-Net algorithm has improved by 2.3%and 5.0%,respectively.