Jumpat
Work in 3D CT medical image segmentation?
Hello,
Thank you for your work. I have a few questions regarding its applicability to the medical domain.
Currently, your evaluation dataset primarily focuses on 3D depth scenes. I attempted to reimplement your code for 3D medical imaging. However, based on your paper and the README description, the "prepare data" step requires the SAM 2D sam_mask and mask_scales. Could you provide guidance on how to create these?
Additionally, if SAM produces suboptimal segmentation results in 2D, is it feasible to use supervised labeling as a replacement? For example, few 3D annotaion label for specific object instead of all.
Thank you for your assistance.
Hi. You can check the two scripts for details.
The extract_segment_everything_masks.py uses the segment everything func of SAM to extract masks for input images one by one. And the get_scale.py use the extracted masks and corresponding 3D-GS model to estimate the 3D scale of the segmented targets.
If you don't want to use SAM for mask extraction, you can just replace the masks generated by SAM with the ones you want to use. You only need to keep the data form the same. And then use get_scale.py is okay.
However, if the label is 3D, you need to render them to be 2D masks, since SAGA does not provide the 3D supervision interface.
