Logit pattern exhibiting repeating rectangle

[hop-soellingeraj]

I am finding several unexpected artifacts in the logit masks returned by sam_vit_h_4b8939.

• exhibiting the pattern of a repeated rectangle (see image 1)
• there are margins at the edge of the image that are anomalously extreme logit values (see image 2)

I am wondering if there is something about the way this model is trained or inferenced that causes this pattern.

Image 1

The repeating rectangle

Image 2

On the left image boundary, there is a 10 pixel wide negative strip that should be positive.

[eyildiz-ugoe]

any update on this? I've experienced this as well