【摘要】 针对行人重识别中因人体姿态多变、遮挡以及外观相似导致识别精度低的问题，提出一种高效的跨空间联合学习方法。首先，通过高效网络EfficientNetV2-S提取行人图像全局特征。其次，设计分支结构，全局分支利用高效多尺度注意力机制与全局平均池化跨空间学习细粒度的全局特征。局部分支使用PCB方法挖掘行人的局部信息，采用特征关系(One-vs.-rest)模块赋予行人不同身体部位空间相关性，缓解相似行人的难辨别问题。最后，拼接分支特征预测行人身份。所提方法在Market-1501和DukeMTMC-ReID两大数据集上的mAP指标分别达到89.2%和79.5%,体现了其先进性。更多还原

【Abstract】 To address the issue of low recognition accuracy in pedestrian re-identification due to variable human postures, occlusions, and similar appearances, an efficient cross-space joint learning method is proposed.Firstly, the global features of pedestrian images are extracted by the efficient network EfficientNetV2-S.Secondly, a branch structure is designed.The global branch uses the efficient multi-scale attention mechanism and global average pooling to learn fine-grained global features across spaces.And the local branch uses the PCB method to mine the local information of pedestrians and the feature relationship(One-vs.-rest) module to endow different body parts of pedestrians with spatial correlation, alleviating the difficulty of identifying similar pedestrians.Finally, the splicing branch features are used to predict the identity of pedestrians.The mAP index of the proposed method on the Market-1501 and DukeMTMC-ReID datasets reaches 89.2% and 79.5% respectively, demonstrating its superiority.更多还原