Hi, I wondered if I can use your codebase to reproduce your best result (IoU of 58.0 on ADE20k) from the SegViTv2 paper, as it seems that the code only covers SegViT v1. I do not need additional v2 features (in particular Shrunk++). Would it be sufficient to exchange the encoder weights (i.e., using BEiTv2 weights)?
I would appreciate any pointers/instructions.
Hi, I wondered if I can use your codebase to reproduce your best result (IoU of 58.0 on ADE20k) from the SegViTv2 paper, as it seems that the code only covers SegViT v1. I do not need additional v2 features (in particular Shrunk++). Would it be sufficient to exchange the encoder weights (i.e., using BEiTv2 weights)?
I would appreciate any pointers/instructions.