Quadruplet-Based Deep Cross-Modal Hashing

Table 2

Comparison to baselines in terms of MAP on two datasets: MIRFLICKR-25K, and Microsoft COCO2014, respectively. The best accuracy is shown in boldface.

Task	Methods		MIRFlickr-25K			MS-COCO
			16bits	32bits	64bits	16bits	32bits	64bits
IT	Handcrafted methods	CMSSH [7]	0.5600	0.5709	0.5836	0.5439	0.5450	0.5410
		SePH [9]	0.6740	0.6813	0.6803	0.4295	0.4353	0.4726
		SCM [8]	0.6354	0.6407	0.6556	0.4252	0.4344	0.4574
		GSPH [10]	0.6068	0.6191	0.6230	0.4427	0.4733	0.4840
	Deep methods	DCMH [13]	0.7316	0.7343	0.7446	0.5228	0.5438	0.5419
		PRDH [14]	0.6952	0.7072	0.7108	0.5238	0.5521	0.5572
		SSAH [15]	0.7745	0.7882	0.7990	0.5127	0.5256	0.5067
		TDH [16]	0.7423	0.7478	0.7512	0.5164	0.5222	0.5276
		DSePH [9]	0.7128	0.7285	0.7422	0.4621	0.4958	0.5112
		QDCMH	0.7635	0.7688	0.7713	0.5286	0.5313	0.5371
TI	Handcrafted methods	CMSSH [7]	0.5726	0.5776	0.5753	0.3793	0.3876	0.3899
		SePH [9]	0.7139	0.7258	0.7294	0.4348	0.4606	0.5195
		SCM [8]	0.6340	0.6458	0.6541	0.4118	0.4183	0.4345
		GSPH [10]	0.6282	0.6458	0.6503	0.5435	0.6039	0.6461
	Deep methods	DCMH [13]	0.7607	0.7737	0.7805	0.4883	0.4942	0.5145
		PRDH [14]	0.7626	0.7718	0.7755	0.5122	0.5190	0.5404
		SSAH [15]	0.7860	0.7974	0.7910	0.4832	0.4831	0.4922
		TDH [16]	0.7516	0.7577	0.7634	0.5198	0.5332	0.5399
		DSePH [9]	0.7422	0.7578	0.7760	0.4616	0.4882	0.5305
		QDCMH	0.7762	0.7725	0.7859	0.5245	0.5398	0.5487