I found a bug in my reimplementation of STELLAR. Precisely the clustering in the new class discovery mechanism should be done for all unlabeld cells not per image. In current implementation new class labels may correspond to clusters with different characteristics. If we want to properly extend our STELLAR implementation to novel class discovery it needs to be adressed.
I found a bug in my reimplementation of STELLAR. Precisely the clustering in the new class discovery mechanism should be done for all unlabeld cells not per image. In current implementation new class labels may correspond to clusters with different characteristics. If we want to properly extend our STELLAR implementation to novel class discovery it needs to be adressed.