Lane Line Detection Based on Object Feature Distillation

In order to meet the real-time requirements of autonomous driving system, existing method directly up-samples encoder’s output feature map pixel-wise prediction, thus neglecting importance decoder for prediction detail features. solve this problem, paper proposes a general lane detection framework based on object distillation. Firstly, with strong ability is added network using direct up-sampling method. Then, in training stage, results generated by are regarded as soft targets through knowledge distillation technology, so that branch can learn more detailed information and have decoder. Finally, stage inference, we only need use instead forward calculation decoder, compared model, it improve performance without additional cost. verify effectiveness framework, applied many mainstream segmentation methods such SCNN, DeepLabv1, ResNet, etc. Experimental show that, under condition no complexity, proposed obtain higher F1Measure CuLane dataset.