Lane Marker Detection Based on Multihead Self-Attention

Table 1

Comparison among different lane mark detection models based on the TuSimple dataset.

Model	Accuracy (%)	FP	FN	FPS
ResNet-18 [31]	92.69	0.0948	0.0822	312
ResNet-34 [31]	92.84	0.0918	0.0796	169
ENet [32]	93.02	0.0886	0.0734	135.4
2-head self-attention (ours)	95.76	0.0407	0.0301	170.2
4-head self-attention (ours)	95.55	0.0339	0.0329	169.5
8-head self-attention (ours)	95.49	0.0414	0.0311	167.5

The significance of bold values means they are the most accurate or they are the lowest error rate.