Which method do you suggest for my segmentation task post-processing step?

Question

I am developing a neural framework for instance segmentation.

At the moment I have quite reasonable outcome (probability map) as follows:

Input image:

Corresponding probability map:

As you may noticed, the outcome has blurred boundaries while I asked the model to highlight DOG location.

My aim is to generate instance borders using:

1: input image information.

2: associated probability map information.

I'm thinking of Graph cut Optimization and Max-Flow/Min-Cut algorithm to extract instance boundaries.

But I wanted to know if you have a more recent exciting suggestion for me?

thanks,