Feature extraction for point clouds

[YunzeMan]

Hi, thanks for your great work. I am trying to use your PonderV2 as a frozen feature extractor for new 3D point clouds, like how people use CLIP/DINOv2 for images.

Could you share some simple scripts or ideas about how we can achieve this with your codebase? This can be very helpful.

[HaoyiZhu]

Hi, thanks for your interest!

You can refer to this function in Ponder model. It will return a sparse_backbone_feat tensor in the data dict. See here.

Alternatively, you can also refer to our downstream finetuning codes, where we load the backbone's weights to the SparseUNet.