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

Table 4

Detection performances (overall mAP and overall AP) of combination methods on test set.

Combination Method		Overall mAP ()			Overall AP ()
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	20.40	15.85	36.15	33.46	23.81	4.75	34.69	17.59	8.08
	Normalized RGB	19.81	15.16	38.10	32.15	22.29	6.06	38.28	11.43	8.65
	Ruta’s RYG	18.07	13.54	33.33	28.58	20.05	2.39	35.30	17.98	4.11
	YCbCr	16.50	12.71	31.31	29.51	15.25	4.67	31.17	14.33	4.07
	HSV	19.70	15.41	37.06	29.23	16.91	6.74	36.00	23.54	5.77
	CIE Lab	17.64	13.31	34.30	26.62	18.27	5.41	34.63	15.82	5.09
Faster R-CNN with Resnet-101	RGB	19.24	14.67	37.91	31.21	20.73	3.79	36.92	14.34	8.44
	Normalized RGB	17.57	13.54	32.86	29.70	18.20	4.87	33.67	11.99	6.98
	Ruta’s RYG	14.72	11.21	28.42	26.55	16.62	4.71	26.27	10.14	4.05
	YCbCr	12.36	9.49	25.02	24.03	10.02	2.83	26.36	8.84	2.05
	HSV	15.76	11.11	32.24	25.07	14.77	5.64	23.06	17.99	8.01
	CIE Lab	10.90	7.63	23.73	19.98	13.79	3.67	20.43	5.28	2.28
R-FCN with Resnet-101	RGB	16.63	11.85	37.27	28.47	13.00	4.92	30.19	18.32	4.85
	Normalized RGB	14.50	10.95	29.97	23.57	14.41	2.50	27.87	14.59	4.08
	Ruta’s RYG	14.21	10.33	26.66	20.89	9.08	3.01	32.75	13.77	5.72
	YCbCr	13.06	9.44	25.05	21.43	10.01	2.50	24.49	14.63	5.28
	HSV	14.66	10.59	29.52	25.40	9.99	3.17	28.39	15.23	5.78
	CIE Lab	12.24	9.06	23.51	14.58	12.43	1.93	27.87	11.80	4.85