Whether the proposed M3SDA method can work on the heterogeneous domain adaptation/generalization task (e.g., VD dataset)?

I am not sure about this question:  whether the proposed M3SDA method can work on the heterogeneous domain adaptation/generalization task (e.g., Visual Decathlon dataset)? Such (mean, variance) alignment operation in M3SDA seems only works for the same-label space setting, is it right? I really want to hear some suggestions/comments from you, thx.