Softmax for Class Probabilities in Panoptic Inference Function

[YilmazKadir]

Thank you for publishing your code. I wanted to ask if there is a small mistake in this line: https://github.com/PRBonn/MaskPLS/blob/f80691232bd25183435fd28dc54026652c209627/mask_pls/models/mask_model.py#LL135C35-L135C35 I believe you should apply a softmax function to get the probabilities similar to what you do in the semantic inference function: https://github.com/PRBonn/MaskPLS/blob/f80691232bd25183435fd28dc54026652c209627/mask_pls/models/mask_model.py#LL119C16-L119C16 Am I missing something or is there a slight mistake?

[rmarcuzzi]

Hi! You can apply the softmax() function and then argmax() to obtain the class probabilities but you can also directly use argmax(). We empirically found that not using the softmax() yielded a slightly better performance. You can try adding it and evaluating and you'll see the difference.