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

Figure 2

The overall framework of our proposed multiscale deep network (M2Net). M2Net is based on ResNet-50 [45] or ResNeXt-101 [46] with supervision from saliency map and object position. M2Net consists of the backbone branch, transformation branch, decoder branch, and centerness prediction branch. The backbone network is used to extract some important saliency features, the transformation part is used to generate robust saliency features, the decoder part is used to generate the final saliency maps, and the last part is used to predict object position in an image.