Perform well on synthetic images but bad on natural images.

[syfbme]

I have done a transfer learning to matting certain objects. Below is my steps: 1: train modnet from scratch. train dataset contains around 60k (6k foreground images and 180 background images) 2: train modnet with soc(30k natural images without synthetic) The model(w/wo soc) performs well on synthetic images but bad on natural images. I think the S branch is is not good enough since the model can't locate the target object well... Maybe the synthetic way is not good enough(I just paste the foreground's target object to background image). Or maybe i should replace the S branch's backbone? Do you have any suggestions? Thanks a lot.

[Tanlang180]

What's your dataset? And I would like to ask where to get the dataset from,haha