search

dengxl0520 / MemSAM

[CVPR 2024 Oral] MemSAM: Taming Segment Anything Model for Echocardiography Video Segmentation.
MIT License
127 stars 11 forks source link

Disscusion about the selection between SAM, MedSAM and SAM-Med2d #4

Closed jinggqu closed 4 months ago

jinggqu commented 4 months ago

Thanks for your valuable work. However, I have several questions about your work:

  1. Why did you choose the original SAM from facebook instead of other variants? As far as I know, there are two SAMs (MedSAM and SAM-Med2D) that are trained on large-scale medical image datasets instead of natural image datasets. Even more, there is also a light-weighted SAM model (LiteMedSAM).
  2. Have you compared the segmentation performance of the above models?
  3. I found that the SAM cannot inference images by batch, how did you solve this problem? (Iterating images one by one?)

Anyway, thank you for open sourcing this novel work.

dengxl0520 commented 4 months ago

Thank you for your attention.

  1. Initially, we chose to use SAMUS for reasons of computing resources, and LiteMedSAM seems to be the later. Moreover, SAMUS is better suited to our task(Echocardiography Segmentation).
  2. For experiments please read our paper.
  3. For MemSAM, inference should be one by one.
jinggqu commented 4 months ago

Thanks for your prompt reply. I will take a look at SAMUS.