PANet + CAPL vs PANet

[alpoler]

Dear Tian,

In application of PANet + CAPL for GFS-Seg, do you follow procedure explained in supplemantary material (CAPL on PANet) ? I wonder whether or not you utilize prototypes to predict y_qry for each class at current episode ( for base classes during inference as well) and compute p_qry with this y_qry. I have also one more question. Do you basically average base prototypes from training set with base classes in support if any ? Since there are no correlation estimator as SCE in standard CAPL, I think that you should average them.

You also mentioned that PANet + CAPL’ differs from PANet only in the training and classifier formation strategies. What kinds of change occur in training different from PANet ?