读论文-SOD-U2Net算法研究

 1 干网讨论

Different from the previous salient object detection models which use backbones (e.g. VGG, ResNet, etc.) as their encoders, our newly proposed U2-Net architecture is backbone free. To validate the backbone free design, we conduct ablation studies on replacing the encoder part of our full size U2-Net with different backbones:

VGG16 and ResNet50.
Practically, we adapt the backbones (VGG-16 and ResNet-50) by adding an extra stage after their last convolutional stages to achieve the same receptive fields with our original U2-Net architecture design.
As shown in Table 2, the models using backbones and our RSUs as decoders achieve better performance than the previous ablations and comparable performance against our small size U2-Net.