Inconsistent segmentation results with pretrained COCO model

[FBehrad]

Hello,

I used the pretrained model on the COCO dataset to predict segmentation masks for the following image. However, I noticed that each time I run the code, the segmentation mask changes. It seems this variation is due to different "hard keep decisions" when retaining or dropping slots.

Given that the model is pretrained, I expected consistent results. Is this behavior normal, or could I be missing something in my implementation?

Thank you for your assistance.