The model we used was U-Net, where the downsampling part is repaced by resnet50.