search

med-air / 3DSAM-adapter

Holistic Adaptation of SAM from 2D to 3D for Promptable Medical Image Segmentation
134 stars 12 forks source link

About pure background images with false positive prompts #19

Open zzzyzh opened 11 months ago

zzzyzh commented 11 months ago

Hi, Thank you for your excellent work at first! I have two questions:

  1. What is your strategy for generating plain background images with false positive hints, and does picking a different background image have any effect?
  2. Since there are false positives, do you randomly select 40 points from the false positives when selecting the points, and why does this help to improve the robustness of the model

Your excellent will be a great help to my research!

peterant330 commented 10 months ago

Hi, Thank you for your excellent work at first! I have two questions:

  1. What is your strategy for generating plain background images with false positive hints, and does picking a different background image have any effect?
  2. Since there are false positives, do you randomly select 40 points from the false positives when selecting the points, and why does this help to improve the robustness of the model

Your excellent will be a great help to my research!

Hi,

I am not sure if I understand your question. For image with pure background, we randomly sample points from the background and use it as prompts during the training phase. And the training process forces the generated background to be blank in spite of the prompts used. For images with foreground, we use both foreground and background points as prompts. The idea is that the attention mechanism can learn to focus on correct prompts from the noisy prompt set so it can help the model be robust towards noisy prompts.