Model did bad on foreground background regions

[Bonnie970]

Hello, I was running your code and the model converged in 9 epochs. But the alpha matte did really bad on the foreground and background region. For example, there are many random regions on the background labelled as foreground. I went through your code, I think the parameter 'wl' in matting_unpool.py already set the training focus on the unknown region. Do you know why the model still struggles with known fg, bg regions?
Thank you!

[protossw512]

Hi, the wl is pixel-wise loss weight, according to original paper, only unknown area have weights, but I found this practice does not work well for segmentation. So I assignment foreground and background small weights, unknown area larger weighs.

If you wanted to do image matting, it does not matter if you have random regions on the background that have positive value, since for image matting, you know these are background and could assign them to 0.
If you wanted to perform segmentation refine, you should process the output of your previous properly, ie, draw unknown area around the boundary of your raw segmentation result.

If you are doing properly, both segmentation refine and matting results should be reasonable, at least not bad.

[Bonnie970]

I see. Thank you very much!