Result
the result is not well,Can you send the results of your implementation?
my loss is not change
I have solved the problem.
Here is my result. Not as good as written in the paper.
Does anyone get the same quality as said in the paper?
122 images are missing the link in alldata_urls.txt file. Did anyone get them from the author? Or do you know how to add tags to improve the quality?
122 images are missing the link in alldata_urls.txt file. Did anyone get them from the author? Or do you know how to add tags to improve the quality?
You can also have a look at this repo to get the remaining images (it's a Chainer implementation) : https://github.com/takiyu/portrait_matting . Don't know if this alone resolve the quality issues yet.