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

Table 5

Comparison of the proposed loss and three other methods on two challenge datasets. M2N: the proposed M2Net with ResNeXt-based backbone CAL: centerness-aware loss; I: Intersection over Union loss [17, 27]; FLoss: F-measure based loss [28]; PPA: pixel position aware loss [19].

Model	DUT-OMRON			PASCAL-S
	Fm	MAE	Sm	Fm	MAE	Sm
M2N + I	0.750	0.048	0.802	0.813	0.078	0.813
M2N + FLoss	0.781	0.051	0.817	0.850	0.074	0.829
M2N + PPA	0.770	0.054	0.841	0.845	0.066	0.857
M2N + CAL	0.791	0.049	0.860	0.858	0.061	0.870