高级检索

面向多种天气场景下目标检测的多域动态平均教师模型

Multi-domain Dynamic Mean Teacher for Object Detection in Complex Weather

  • 摘要: 现有基于深度学习的目标检测模型由于复杂天气,使得现有方法的结果大幅降低, 为了有效地消除不同天气场景带来的域差异问题, 提出一种多域动态平均教师模型. 首先, 引入多域平均教师模块, 为多个不同天气场景下目标域数据生成伪标签; 然后, 引入基于学生网络的风格迁移模块, 解决多域任务中学生网络对不同目标域泛化能力差的问题, 可有效地减小源域与不同目标域之间的差异, 提升学生网络对不同目标域的泛化能力; 最后, 提出基于教师网络的动态过滤伪标签模块, 根据教师网络对不同目标域的学习效果动态调整过滤伪标签的阈值, 提升每个目标域伪标签质量. 在FoggyCityscapes & RainCityscapes和Dusk-rain & Night-rain数据集上的实验结果表明, 所提模型分别获得了40.3%和31.4%的精度, 在雨天, 雾天和夜晚等多种复杂天气场景下都优于其他对比方法.

     

    Abstract: The performance of existing deep learning-based object detection models significantly degrades due to the influence of complex weather. To effectively eliminate the problem of domain differences caused by different weather scenes, propose a multi-domain dynamic mean teacher model. First, introduce a multi-domain mean teacher module to generate pseudo-labels for target domain data in multiple different weather scenes. Then, introduce a student network-based style transfer module to solve the problem of weak generalization ability of student network to different target domains in multi-domain tasks. It can effectively reduce the difference between source domain and different target domains and improve the generalization ability of the student network to different target domains. Finally, a teacher network-based dynamic filtering pseudo-label module is used to dynamically adjust the threshold values of filtering pseudo labels according to the learning effect of the teacher network on different target domains and improve the quality of pseudo labels for each target domain. Experiments on FoggyCityscapes & RainCityscapes and Dusk-rain & Night-rain datasets show that the proposed model achieves 40.3% and 31.4% accuracy respectively, and outperforms other comparison methods in complex weather scenes such as rain, fog and night.

     

/

返回文章
返回