YazhouZhu19
commented
1 day ago Hi, thanks for your attention,
The Supervoxel algorithm is based on the cluster. The Supervoxel method and our model don't know which semantic information each cls_idx belongs to. The cis_idx are just simple assigned number for each clustered region.
In the Setting2, the excluding operation is performed in the slice dimension, i.e., the 2D slices which contains the chosen classes will be excluded. And the Supervoxel algorithm does not change the dimension of slices.
Dear author,
Thank you for your excellent work and the detailed documentation, which helped me quickly run the code. However, I have some questions regarding the use of Supervoxels as a supervisory signal (or label), and I hope you can help me with them!
I noticed that in the CHAOST2 dataset, when using Supervoxels, the number of cls_idx obtained ([0,1,…,35,…]) is much larger than when using GT as a supervisory signal. I am wondering if such labels still carry the same semantic information as the original labels, such as “liver,” “right kidney,” “left kidney,” “spleen”?
If they do, could you kindly explain what each class represented by the cls_idx in the Supervoxels corresponds to in terms of specific categories?
If the Supervoxels are derived from clustering and do not contain semantic information, why is it possible during training in setting2 to exclude the four categories “liver,” “right kidney,” “left kidney,” and “spleen” by excluding cls_idx = [1,2,3,4]?
If my understanding is incorrect, please feel free to correct me.
Thank you again for your help!