SOTA Panoptic Methods Implementation

[aselimc]

Hello,

I see that #1059 #1044 #1132 has already mentioned implementation of Mask2Former/EoMT; however, it seems like there are no progress on them. I see that you have concerns regarding these models being panoptic models which would require (maybe?) changes in the fundamental code. However, I wanted to ask if you have plans to implement these models by yourself or you'd just wait and see if someone else can come-up with an elegant implementation that could work?

The following are some models, that I could found are currently still the SOTA models,

Mask2Former
OneFormer
EoMT
kMaX-DeepLab

Also there is another question that I have is that DPT only works with timm encoders. Do you plan to support native encoders as well?

Best