Question about HQ-Output Token and weight updates in the frozen Mask Decoder

[Linn0910]

Hello,

Thank you for your great work on the HQ-SAM model! I have a question regarding the role of the HQ-Output Token in the model when interacting with the frozen Mask Decoder.
From the architecture diagram, I understand that the HQ-Output Token is integrated into the frozen Mask Decoder to improve segmentation accuracy. However, I am curious about how the HQ-Output Token's weights are updated during training, given that the Mask Decoder itself is frozen and its weights are not updated.
Here are my specific questions:

1.Since the Mask Decoder is frozen, how are the HQ-Output Token's weights updated during training?
2.Does the HQ-Output Token rely solely on the Global-local Fusion and MLP layers for weight updates, or does it interact with the Mask Decoder in a different way for updates?
3.How does the error correction mechanism contribute to the HQ-Output Token’s learning in this setup?
I would greatly appreciate it if you could clarify these points. Thank you again for your time and for sharing your amazing research!

Best regards,
Lin