Lightweight Lane Marking Detection CNNs by Self Soft Label Attention

Xuefeng Yang,Yanxun Yu,Zhen Zhang,Yu Huang,Zhihui Liu,Zhongbin Niu,Hongwei Chai,Chenglu Wu,Zhijiang Du,Song Li

DOI: https://doi.org/10.1007/s11042-022-13442-6

IF: 2.577

2022-01-01

Multimedia Tools and Applications

Abstract:Lane detection models based deep semantic segmentation often fail in challenging scenarios for lacking enough contextual information. In this paper, we first point out that a soft label extracted from a trained semantic segmentation model is capable of encoding rich contextual information. Then, we find that introducing the soft label extracted from the top layer of a well-trained model to low-level features gains substantial improvement. Based on this observation, we propose a novel framework, i.e., Soft Label Attention(SLA), which allows a model to learn contextual attention from the higher layer through a loss function. Significantly, our SLA does not increase the run time in the inference stage because it is only applied in the training stage. We apply the proposed SLA to modified ERFNet and propose an attention deep neural network (DNN) for lane detection (ERFNet-SLA). Tests on the TuSimple dataset show that the ERFNet-SLA outperforms ResNet-18 and ResNet-34 by 3.4% and 3.3% respectively with at least 2.3× faster in computation speed, and is 12.5× faster than SCNN with 0.6% performance loss. Tests on the CULane dataset show that the ERFNet-SLA outperforms ResNet-50 by 5.6% and is 16× faster than ResNet-101 with 0.5% performance loss.