Advances in Multimedia

Research Article

Multiscale Deep Network with Centerness-Aware Loss for Salient Object Detection

Table 2

Performance comparison with state-of-the-art methods on five popular saliency datasets. MAE (smaller is better), max F-measure (larger is better), and E-measure (larger is better) are used to measure the model performance.


Method	DUT-OMRON			DUTS			ECSSD			PASCAL-S			HKU-IS
Method	F_m	MAE	S_m	F_m	MAE	S_m	F_m	MAE	S_m	F_m	MAE	S_m	F_m	MAE	S_m

VGG backbone
Amulet₁₇	0.647	0.098	0.781	0.678	0.085	0.804	0.868	0.059	0.894	0.769	0.099	0.819	0.841	0.051	0.886
NLDF₁₇	0.684	0.080	0.770	—	—	—	0.878	0.063	0.875	0.780	0.101	0.801	0.874	0.048	0.879
BMPM₁₈	0.692	0.064	0.809	0.745	0.049	0.862	0.868	0.045	0.911	0.771	0.075	0.845	0.871	0.039	0.907
C2SNet₁₈	0.683	0.072	0.798	0.716	0.063	0.828	0.864	0.055	0.893	0.769	0.083	0.835	0.851	0.048	0.883
RAS₁₈	0.713	0.062	0.814	0.751	0.059	0.839	0.889	0.056	0.893	0.785	0.106	0.793	0.871	0.045	0.887
PiCANet₁₈	0.710	0.068	0.826	0.749	0.054	0.861	0.885	0.046	0.914	0.801	0.079	0.849	0.870	0.042	0.906
CPD₁₉	0.745	0.057	0.818	0.813	0.043	0.867	0.915	0.040	0.910	0.830	0.075	0.841	0.896	0.033	0.904
MINet₂₀	0.741	0.057	0.821	0.823	0.039	0.875	0.922	0.036	0.919	0.840	0.066	0.852	0.904	0.031	0.912

ResNet backbone
DGRL₁₈	0.733	0.062	0.806	0.794	0.050	0.842	0.906	0.041	0.903	0.827	0.073	0.837	0.890	0.036	0.894
PiCANet₁₈	0.717	0.065	0.832	0.759	0.051	0.869	0.886	0.046	0.917	0.802	0.078	0.852	0.870	0.043	0.904
BASNet₁₉	0.756	0.057	0.836	0.791	0.048	0.866	0.880	0.037	0.916	0.777	0.079	0.834	0.896	0.032	0.909
EGNet₁₉	0.756	0.053	0.841	0.815	0.039	0.887	0.920	0.037	0.925	0.829	0.076	0.850	0.901	0.031	0.918
PoolNet₁₉	0.747	0.056	0.836	0.809	0.040	0.883	0.915	0.039	0.921	0.828	0.076	0.849	0.899	0.032	0.917
CPD₁₉	0.747	0.056	0.825	0.805	0.043	0.869	0.917	0.037	0.918	0.829	0.074	0.844	0.891	0.034	0.906
SCRN₁₉	0.746	0.056	0.837	0.809	0.040	0.885	0.918	0.038	0.927	0.837	0.066	0.865	0.897	0.034	0.916
MINet₂₀	0.755	0.055	0.833	0.828	0.037	0.884	0.925	0.034	0.925	0.840	0.066	0.854	0.909	0.029	0.919
GateNet₂₀	0.746	0.055	0.838	0.807	0.040	0.885	0.916	0.040	0.920	0.830	0.071	0.854	0.899	0.033	0.915
Ours	0.774	0.054	0.846	0.837	0.038	0.889	0.926	0.033	0.927	0.847	0.064	0.863	0.913	0.029	0.922

ResNeXt backbone
R3Net₁₈	0.747	0.062	0.815	—	—	—	0.914	0.040	0.910	0.803	0.095	0.803	0.894	0.036	0.895
GateNet₂₀	0.762	0.051	0.849	0.816	0.035	0.897	0.917	0.035	0.929	0.827	0.065	0.865	0.903	0.030	0.925
Ours	0.791	0.049	0.860	0.857	0.034	0.900	0.934	0.032	0.933	0.858	0.061	0.870	0.923	0.025	0.929

Bold, italics, and underline indicate the best, second best, and third best performance. “—” means that the author has not provided corresponding saliency maps.