Lans1ng/PointSAM

the shape of the predicted instance feature

Closed this issue · 2 comments

Hi, I’m interested in your great work and have a question.

In the Methodology section, part C "Memory Bank Updating," you mention: “Given the predicted instance feature f^p_k​ with a positive prompt from the teacher’s image encoder.”

Is the shape of the predicted instance feature f^p_k the same as f^t_k and extracted from the location of a point prompt? Or is it calculated from the features in the instance mask?

Thanks for your attention. The instance feature of the teacher encoder is also extracted from the location of a point prompt. Of course, extracting features using the predicted mask and taking the average is also worth trying.

Thanks for your reply!