Implementation of PRM module

[wds1998]

Hello, your work MGMatting is very excellent. I read your paper and code, I found that the PRM module code in the paper is implemented using two-layer convolution. Do you have any code for PRM in the paper? Does PRM bring a great gain effect when you make experiments, especially for the details such as the hairline of the portrait matting? Thank you very much for answering this question.

[yucornetto]

Hi, please refer to the decoder file in the code-base, where additional convs of PRM are defined. But the usage of PRM also include some code parts in utils.py and main.py, in terms of how to use later-stage prediction to refine the previous stage's one, and also related loss computation.

[wds1998]

Hi, please refer to the decoder file in the code-base, where additional convs of PRM are defined. But the usage of PRM also include some code parts in utils.py and main.py, in terms of how to use later-stage prediction to refine the previous stage's one, and also related loss computation.

Thank you very much.In addition, I would like to ask if you encountered a situation where the background pixel was incorrectly predicted as the foreground pixel during the test (the value of the background pixel is predicted to be a low confidence value), do you know how to solve it or how to strengthen the guidance of mask for the final alpha result.