Some image patches have the binary mask inversed

[pagraf]

Hi, thanks for sharing your work. By applying to some images, i noticed that in some patches in the predicted mask, black and white is inversed, compared to the majority of the prediction. Do you know why? Have you solved it? It is apparent also in the image you shared. My prediction:

[aliaksandr960]

Hi, thanks. When I tried to reproduce a similar issue and look at the results I did not find that the inference image is inverted, but I found that SAM sometimes segments background and sometimes not. It looks like the instability of SAM NN.

[aliaksandr960]

If your results are really inverted, please provide me images or code to reproduce it. Unfortunately, your screenshots are too small to see all the details.

But anyway, such behavior looks ugly, and it is good idea to fix it.

[pagraf]

It happens with any image. However, i somehow solved it by replacing line 55 in sameo.py with resulting_mask = np.ones((h, w), dtype=np.uint8). I changed the np.zeros to np.ones.

[hectorcarrion]

It happens with any image. However, i somehow solved it by replacing line 55 in sameo.py with resulting_mask = np.ones((h, w), dtype=np.uint8). I changed the np.zeros to np.ones.

This works great! Thanks for providing an easy fix.