How exactly are the predictions made for ignore_label handled in the panoptic deeplab model? about deeplab2

[csrhddlam]

Hi,

Thanks for raising the issue and checking all the details. As far as I know, there is no explicit ignore_label handling in the panoptic-deeplab post-processing. I think in the current implementation, the training "implicitly and automatically" learns not to predict ignore_label, since there is not any positive label for the ignore_label class. That said, in a sufficiently trained model, the current implementation will not be an issue.

If you want to be really rigorous, a simple diagnose could be done by setting a large positive or negative bias towards the ignore_label class in a trained (downloaded) model and checking the visualizations and results. But again, I don't think explicitly dealing with the ignore_label will result in any performance difference.

[free-bit]

Thank you very much for the explanation. It is very good to learn these details for me.