zero-shot accuracy on scannet200

[seanzhuh]

Hi, have you ever test the zero-shot accuracy on scannet200, i.e., replace those class-agnostic mask proposals predicted by Mask3D and use ground-truth instances as input to your mask feature computation module to get the mask features, then dot product with text embeddings from CLIP text encoder and select the maximum index as the predicted label for ground-truths? I just wonder how accurate CLIP is.

[elisabettafedele]

Hi, thank you for the interest in our work!

Yours is a very good question. We actually experimented this and put our results in the main paper in Table 5.

Let us know if you have any other question.

Best,
Elisabetta