How to run text-prompt model on FSCD-LVIS?

[Yiming-M]

Hi @jerpelhan

Thanks for the great work!

In your paper, it seems that prompt-based counting is only conducted on FSC-147, as this dataset explicitly provides the name of the class of interest in each image.
I'm wondering, is there any possibility to do something similar on FSCD-LVIS?

Looking forward to your reply.

Many thanks,
Yiming

[jerpelhan]

Sorry for the late reply. I have not tried text-prompt-based counting on LVIS. The FSCD-147 dataset is more simple since it typically contains one or a few classes per image. Most state-of-the-art zero-shot counting methods tend to count all objects in an image. We use this as an advantage in text-prompt based counting. DAVE runs zero-shot model and then filters candidates with the text-specified class.