How to evaluate zero-shot semantic segmentation?

[Pixie8888]

Hi,

In the paper, only generalized zero-shot setting is conducted in the semantic segmentation. I wonder how to evaluate zero-shot segmentation performance, ie only test the performance of segmenting novel classes on the testing point cloud. I imagined one potential way: the final classifier is only trained on generated unseen class feature, so the classifier can only differentiate different new classes. However, the testing point cloud consists both seen and unseen classes. How can the classifier to differentiate seen classes from unseen classes?