I am trying to understand the technique used in Vista3D for maintaining segmentation mask consistency between the slices of the 3D volume. As I understood from the Algorithm 2 on supervoxel generation in the paper is that the authors used SLIC over SAMPredictor features output. But I am unsure about the upsampling part. Is this achieved from learning a CNN or just interpolation.

Specifically, line 142 in the code below has this function get_feature_upsampled(). I want to know it's implementation details to really understand the process better and how and why SLIC works.

https://github.com/Project-MONAI/VISTA/blob/main/vista3d/scripts/slic_process_sam.py

Thanks in advance!