A Knowledge Distillation Network Combining Adversarial Training and Intermediate Feature Extraction for Lane Line Detection

Abstract

Lane line detection is an important input of the automatic driving system and the assisted driving system. It is deployed on the vehicle end, with limited resources and high requirements for real-time performance and detection accuracy. We explore a new knowledge distillation method for lane line detection, in which the student network can acquire knowledge not only from the output features of the teacher network but also from the intermediate process of the teacher network. The knowledge distillation in the intermediate process named important feature correlations distillation compares the correlation between the feature maps of the teacher network and the student network. The knowledge distillation of the output results named semantic consistency distillation allows the student network to learn the output feature knowledge of the teacher network by integrating confrontation training into the knowledge distillation method. Experimental results demonstrate that our knowledge distillation method works well and light models can benefit from the distillation method.