search

bowang-lab / MedSAM

Segment Anything in Medical Images
https://www.nature.com/articles/s41467-024-44824-z
Apache License 2.0
2.48k stars 323 forks source link

About the bounding box #280

Closed midcola closed 3 days ago

midcola commented 1 week ago

Hi there, thx for sharing your work first!

I've met some problem with the bounding box:

  1. How do you decide bounding boxes when evaluating?
  2. How do you decide bounding boxes of samples in training data, of which mask is all zero (negative sample)?

Thank you for your reply in advance!

JunMa11 commented 1 week ago

Hi @midcola ,

  1. It was generated from ground truth.
  2. Only the foreground masks were used during training.
midcola commented 1 week ago

Hi @midcola ,

  1. It was generated from ground truth.
  2. Only the foreground masks were used during training.

So that means, we will tell the model, to some degree, the ground truth, like what we do in training?

JunMa11 commented 4 days ago

No. During inference, it only inputs the image and bounding box, which are not ground truth.