Simple 3D Mask-RCNN Implementation

Hello everyone,

I am wondering if there is a simple 3D Mask-RCNN code? Also, is there trained models that I can simply just import like ResNet. Please note that I need 3D not 2D.

Any other state-of-the-art 3D semantic segmentation/Instance segmentation models?

Thank you very much.

Hi Rick!

I am not aware of any pre-packaged (or pre-trained) 3D Mask-RCNN

A search will lead you to a number of pytorch 3D U-Net implementations.
(I don’t know which, if any, of them are any good.)

Be aware that U-Net performs semantic (not instance) segmentation.
(People do get good instance segmentation from U-Net in some use
cases by post-processing (e.g., running connected components) the
results of U-Net’s semantic segmentation.)

Note, for 3D images of any size, any such 3D model will quickly become
resource hungry, both in terms of memory and computation time.


K. Frank

Hi Frank,

Thank you very much. I have tried U-nets for Semantic segmentation and now I am looking for something different to U-net to see if it would make any difference. And you are absolutely correct about resource problem, it is the most difficult part dealing with 3D-Unets for me.