An Efficient Color Space for Deep-Learning Based Traffic Light Recognition

Table 5

Detection performances ([email protected] and [email protected]) of combination methods on test set.

Combination Method		[email protected] ()			[email protected] ()
Ensemble Network Model	Color Space	total	small	non small	green	red	yellow	red left	green left	off
Faster R-CNN with Inception-Resnet-v2	RGB	38.48	31.27	57.79	70.56	52.12	8.49	59.11	27.13	13.44
	Normalized RGB	38.24	31.42	59.87	70.43	52.09	10.98	63.94	17.39	14.60
	Ruta’s RYG	35.94	29.16	52.99	65.02	49.77	06.03	57.87	28.76	8.16
	YCbCr	35.55	29.32	51.83	68.68	41.30	9.53	58.91	26.07	8.83
	HSV	35.13	28.82	56.76	58.55	38.50	12.94	57.89	32.88	10.04
	CIE Lab	32.19	25.45	53.05	54.26	41.84	8.47	55.71	24.00	8.84
Faster R-CNN with Resnet-101	RGB	37.24	30.25	61.45	65.23	47.68	6.82	63.11	24.37	16.23
	Normalized RGB	34.24	28.32	51.82	64.11	43.46	8.20	57.30	19.63	12.72
	Ruta’s RYG	31.96	26.14	50.63	61.55	41.21	13.01	50.04	18.11	7.85
	YCbCr	26.17	21.44	42.64	56.82	27.16	5.52	47.45	15.57	4.48
	HSV	30.30	22.69	54.00	52.50	34.45	11.02	41.49	27.88	14.44
	CIE Lab	24.71	18.86	41.38	46.99	33.59	6.56	46.18	9.48	5.48
R-FCN with Resnet-101	RGB	34.88	27.33	62.19	64.76	36.48	10.07	55.44	30.93	11.57
	Normalized RGB	32.16	26.18	52.80	58.86	38.17	5.99	54.03	26.46	9.43
	Ruta’s RYG	31.21	24.78	47.20	55.03	30.14	7.13	60.38	23.28	11.29
	YCbCr	30.42	23.18	50.18	57.18	29.10	5.45	49.49	30.42	10.87
	HSV	30.05	23.25	51.61	56.33	28.48	7.58	50.58	26.10	11.25
	CIE Lab	27.33	21.63	44.38	46.86	32.74	4.41	48.95	21.54	9.50