search

ZrrSkywalker / PointCLIP

[CVPR 2022] PointCLIP: Point Cloud Understanding by CLIP
335 stars 31 forks source link

Zero Learning Question #4

Closed DuanhaoranCC closed 2 years ago

DuanhaoranCC commented 2 years ago

Hello, Zero Learning is to make predictions on categories you don't see. CLIP self supervised training on 400 million picture text pairs will inevitably see categories you don't see. Is this also called Zero Learning? Thank you for your patience.

ZrrSkywalker commented 2 years ago

Thanks for the question. We inherit the same zero-shot/few-shot setup following the original CLIP paper, which is, predicting on a new dataset without any in-dataset training. It differs from the traditional query/support set division and is more challenging. Traditional setup trains and tests within the same dataset of different categories, but CLIP and our PointCLIP test on a completely new dataset without considering the categories.