CCAH: A CLIP-Based Cycle Alignment Hashing Method for Unsupervised Vision-Text Retrieval

Table 4

MAP@50 results at MIRFlickr-25K and NUS-WIDE for ablation analysis.

Method	Firlickr-25K								NUS-WIDE
	I->T				T->I				I->T				T->I
Bits	16	32	64	128	16	32	64	128	16	32	64	128	16	32	64	128
CCAH	0.863	0.879	0.899	0.91	0.891	0.908	0.914	0.913	0.715	0.754	0.775	0.787	0.834	0.851	0.864	0.874
GAT	0.892	0.922	0.935	0.905	0.886	0.886	0.897	0.896	0.796	0.829	0.841	0.85	0.774	0.792	0.8	0.808
CLIP	0.869	0.871	0.888	0.901	0.89	0.895	0.906	0.908	0.735	0.759	0.781	0.793	0.836	0.842	0.858	0.869
CA	0.859	0.876	0.891	0.907	0.883	0.899	0.906	0.904	0.713	0.748	0.771	0.784	0.828	0.848	0.861	0.87
ALL	0.863	0.877	0.895	0.903	0.846	0.86	0.881	0.882	0.775	0.805	0.818	0.827	0.772	0.791	0.804	0.815