SAM model returns inversed

[francois-dibulo]

Describe the bug

I tried using the SAM model to remove the background and normally I get this result (using e.g. u2netp)

With SAM I am getting the opposite:

To Reproduce Code:

  session = new_session("sam")
  sam_prompt = [{"type": "rectangle", "data": [0, 0, width, height], "label": 1}]
  res = remove(cvImage, session=session, sam_prompt=sam_prompt)

Input image

Expected behavior

• Any idea why SAM returns the opposite?
• Where can I find more information on e.g. what the label property is doing?
• For those kind of images, which model is the best?

Thank you!

Rembg version: v2.0.57