search

charusharma1991 / SSL_PointClouds

34 stars 5 forks source link

About the splits of dataset in FSL settiing. #5

Open gaunzi opened 3 years ago

gaunzi commented 3 years ago

Hi Charu,

Thanks for sharing this interesting work. May I know more details about the few-shot settings defined in this work? The FSL settings mentioned in the paper seem to be different from the general 2D FSL settings, where the classes (base-class) of training data are disjoint from the classes of (novel-class) testing data. While the examples used for training and testing in this work seemingly come from the same K' classes.

Looking forward to your reply. Thanks again!

gaunzi commented 3 years ago

Hi Charu,

By the way, could you share the splits (.pkl files) of ModelNet40?

Thanks a lot!

charusharma1991 commented 3 years ago

Hi Gaunzi,

Thanks for your interest in our work! There are different setups defined by different work for FSL. However, we follow setting similar to the one in "Few-shot learning with graph neural networks".

Yes, I can provide splits for MN40 although you can also generate your own by randomly sampling m PCs for each of the randomly sampled K' classes from K classes. Thanks!