基于多阶段知识蒸馏的行人再识别

Person Re-identification based on Multi-stage Knowledge Distillation

行人再识别模型在多目标行人跟踪、行人抓拍去重等应用场景下具有较高应用价值。在较多使用场景中对模型推理速度提出了较高要求,通常行人再识别模型在保证推理速度大幅提高的同时,模型的识别准确率却大幅降低。当前的知识蒸馏方法大多针对分类任务、目标检测任务、分割任务等,而对于行人再识别准确率最为关键的行人特征部分并未做出重点优化。针对模型提取的特征进行蒸馏,模型骨干网络的多个阶段也同时进行辅助蒸馏,将骨干网络为ResNet18的PCB网络的识别准确率提高到与ResNet50相当的水平,相比于蒸馏之前在Market1501、DukeMTMCreID、CUHK03数据集上学生网络ResNet18的准确度分别提高了4.21%、5.8%、5.82%。

Person re-identification model has high application value in application scenarios such as multi-target person tracking and person capture to remove duplicates.In many practical scenarios,higher requirements are put forward for the model’s reasoning speed.Usually,the person re-identification model guarantees a significant increase in the reasoning speed while the model’s recognition accuracy is greatly reduced.The current knowledge distillation methods are mostly aimed at classification tasks,target detection tasks,segmentation tasks,etc.And the key features of person that are most critical to the accuracy of person re-recognition are not optimized.Distillation is performed on the features extracted by the model,and multiple stages of the model backbone network are also simultaneously distilled to improve the recognition accuracy of the PCB network whose backbone network is ResNet18 to a level equivalent to that of ResNet50.Compared with the previous distillation in Market1501,DukeMTMC-reID,CUHK03 datasets,the accuracy of the student network ResNet18 increased by 4.21%,5.8%,and 5.82%respectively.