Why not directly use models from Mask2Former in the MODEL_ZOO.md?

[yxchng]

From what I understood, the only difference between mask proposal network in this repo is the number of classes predicted (binary vs N dataset classes) and N predictions can be converted to binary predictions. Are there any other difference? Just curious, have you done any ablation to verify if binary prediction is necessarily better?

[zh-ding]

The models from Mask2Former are mask-aware as they utilize a classification head, which would make the mask proposal network biased toward the base classes causing the performance of the novel classes to be lower. The class-agnostic mask proposal network, on the other hand, learns to output more general object/stuff masks without the awareness of each specific class information and thus would be more suitable for open-vocabulary image segmentation. Hope this helps!