Table 2.

Recognition rates of proposed method compared with the state-of-the-art approaches on (top) UIUC dataset and (bottom) UMD dataset. N_t is the number of training images in each class.

Nt	VG [30]	MFS [33]	Lazebnik [16]	Zhang [35]	SIFT [18]	SURF [2]	DAISY [31]	ORB [25]	CARD [1]	MROGH [10]	Our method
5	82.86	82.24	91.12	88.62	91.96	90.73	86.80	79.03	73.99	88.76	90.86
10	87.85	88.36	94.42	93.17	95.42	95.15	92.54	86.26	83.00	94.13	95.55
15	90.62	91.38	96.64	95.33	96.87	96.14	94.16	89.40	87.18	95.93	97.07
20	92.31	92.74	97.02	96.67	97.84	96.75	95.21	90.73	89.69	96.82	97.91

Nt	VG [30]	MFS [33]	Lazebnik [16]	Zhang [35]	SIFT [18]	SURF [2]	DAISY [31]	ORB [25]	CARD [1]	MROGH [10]	Our method
5	90.92	85.63	90.71	91.56	91.68	90.41	90.81	81.85	84.38	90.39	91.23
10	94.09	90.82	94.54	96.00	96.01	94.49	94.92	87.87	90.41	94.54	96.06
15	96.22	92.67	96.29	96.79	97.21	96.13	96.47	90.87	93.05	96.01	97.59
20	96.36	93.93	96.95	97.62	97.64	96.98	97.58	92.84	94.23	97.03	98.20