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

Table 4

Ablation analysis on two challenge datasets. FPN-1: FPN [48] with ResNet-50 backbone; FPN-2: FPN [48] with ResNeXt-101 backbone; M2N-1: the proposed M2Net with ResNet-based backbone; M2N-2: the proposed M2Net with ResNeXt-based backbone B: binary cross entropy; C: centerness; I: Intersection over Union.

Model	Loss	DUT-OMRON			PASCAL-S
		Fm	MAE	Sm	Fm	MAE	Sm
FPN-1	+B	0.711	0.064	0.815	0.814	0.074	0.849
M2N-1	+B	0.742	0.055	0.840	0.830	0.069	0.861
M2N-1	+B + C	0.748	0.055	0.842	0.836	0.067	0.862
M2N-1	+B + C + I	0.774	0.054	0.846	0.847	0.064	0.863
FPN-2	+B	0.745	0.056	0.842	0.835	0.067	0.861
M2N-2	+B	0.761	0.051	0.851	0.845	0.063	0.864
M2N-2	+B + C	0.768	0.050	0.855	0.852	0.061	0.866
M2N-2	+B + C + I	0.791	0.049	0.860	0.858	0.061	0.870