Image Border about rembg

[jjangga0214]

@danielgatis

Hi,
Try this:
https://github.com/danielgatis/rembg#advance-usage

This does not exist anymore.
Is it moved to https://github.com/danielgatis/rembg/blob/main/USAGE.md?
And was what you especially suggest alpha metting?
Thanks.

[danielgatis]

Hi,
Try this:
https://github.com/danielgatis/rembg#advance-usage

[tomstevelfq]

Hi,ashkaps,
Have you found the reason why the pixel degradation on the contours after alpha matting?

[alfonsmartinez]

works perfectly in this case.

thank you very much
good work

[ashkaps]

Hello @danielgatis thank you for the awesome work. This is great.
While I was trying this out I struggled with a few issues initially where I would see a fringe on the border of the object.
After your implementation of alpha matting, this issue got resolved but I am starting to see pixel degradation on the contours.
I was wondering how you ended up with 240 - 10 -10? I have tried a few different combinations and each fix one set of images and break the other set :D.
In your opinion what is the best way to handle all varied set of images?

Images - https://drive.google.com/file/d/1CK_pIsH9jX09xC4IyIF8Jkc0XclL3fju/view?usp=sharing

[ashkaps]

hey @danielgatis thanks for reaching out to the pymatting team. Seeing their response there really isn't a general way to predict the boundaries of an object even with a neural network.
Considering a neural network will be trained on a particular usecase.
I have been checking out results from a site called remove.bg and their results look good in most cases. Appears that they are somehow deleting the boundary pixels or more so they have a crisp different in background and foreground. They are actually chipping away a minute portion of the image. Thoughts?

[mingkang1993]

嗨，ashkaps， 您找到 alpha matting 后轮廓上的像素退化的原因了吗？

epsilon=1e-5, Try this