基于分组卷积的多尺度注意力行人重识别方法

Multi-scale attention person re-identification method based on group convolution

针对传统基于注意力的行人重识别方法存在仅学习全局注意力图以及对明显不同领域或体系结构无法适应参数调整等问题,提出使用ResNet-50作为基础模型。通过结合多尺度注意力机制和分组卷积,实现行人重识别。此外,采用分类损失和三元组损失,使网络模型能够学习具有鉴别性的特征。实验结果表明,所提出的基于分组卷积的多尺度注意力机制的行人重识别模型在性能上取得了提升,同时增强了网络模型在未知数据域的泛化能力,并加快了模型的运算速度。

In response to the problems of traditional attention based pedestrian re recognition methods only learning global attention maps and being unable to adapt to parameter adjustments in significantly different domains or architectures,ResNet-50 is proposed as the basic model.By combining multi-scale attention mechanism and grouped convolution,pedestrian re-identification is achieved.In addition,the use of classification loss and triplet loss enables the network model to learn discriminative features.The experimental results show that the proposed pedestrian re identification model based on group convolution and multi-scale attention mechanism has achieved performance improvement,while enhancing the network models generalization ability in unknown data domains and speeding up the models computation speed.