Clarification on evaluation metric M-IoU

[roysubhankar]

Hi Fabio,

I would like to ask how do you compute the final M-IoU in the tables. In the table below I see that to compute the Old and All M-IoU for WILSON you do the following:

1. To compute Old you take a mean of the old classes (excluding the background class)
2. To compute All you take a mean of all the classes (including the background class)

If I recall well, in your previous work MiB you said somewhere that you do not take into account the background classes while computing the final metric as bkg is not a class of interest. However, in this work you consider the bkg class to compute the All M-IoU. Could you please clarify which calculation should be adopted?

[fcdl94]

Ciao @roysubhankar,

Sorry for the late answer. Actually, I started using the background class after the last Continual SemSeg papers, that are using the background for the All class. However, the results are in line, but nearly 1 point less when excluding the background.

Thanks for opening the issue, I should clarifty it better.