How to train custom models about rembg HOT 2 CLOSED

I found one amazing thing which helps a lot with background removal.

staring image is:

and result is:

which is not great but when I add border to image by:
convert -border 100x100 -bordercolor white input.jpg inputwithborder.png

and then run rembg i get:

which is completely acceptable.

Any thoughts what will happen when I will send to rembg all images with added white border?

I tested it with other different images and still getting nice results. But it will complicate way how to process images, it would be amazing if it this addition of border could be done directly in rembg. What do you think?

from rembg.

Hi @ondrejbartas ,

For most cases your predefined models work great. But in some cases I would like to somehow train/update model which is processing pictures.

Under the hood, this project uses a neural network called u2net. You can get more information about training a new model in the project repo.

I tested it with other different images and still getting nice results. But it will complicate way how to process images, it would be amazing if it this addition of border could be done directly in rembg. What do you think?

I think this step should be done outside of rembg, because it is not sure this pre-processing step will work for any image.

from rembg.