Lane Marker Detection Based on Multihead Self-Attention

Table 2

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

Model	Total	Normal	Crowd	Highlight	Shadow	Arrow	Curve	Cross	Night	No line	FPS
SCNN [17]	71.60	90.60	69.70	58.50	66.90	84.10	64.40	1990	66.10	43.40	7.5
ERF-Net [29]	73.10	91.50	71.60	66.01	71.30	87.20	71.60	2199	67.10	45.10	85.87
R-34-SAD [28]	70.70	89.90	68.50	59.90	67.70	83.80	66.02	1960	64.60	42.20	75
R-34-E2E [30]	71.50	90.40	69.90	61.50	68.10	83.70	69.80	2077	63.20	45.01	—
2-head self-attention (ours)	75.43	91.46	73.62	66.24	64.07	87.09	66.19	1329	69.96	48.68	170.5
4-head self-attention (ours)	75.52	91.34	73.56	66.18	66.81	86.79	65.60	1115	70.30	47.89	169.6
8-head self-attention (ours)	75.55	91.43	73.85	66.19	69.68	87.02	65.81	1286	69.85	48.18	167.8

The significance of bold values means that F1 is the most highest one.