search

Pointcept / PointTransformerV2

[NeurIPS'22] An official PyTorch implementation of PTv2.
357 stars 25 forks source link

About Patch Partition #48

Closed Litoda closed 3 months ago

Litoda commented 3 months ago

Hello, may I ask, if the partition-based pooling/ patch division only happens at the beginning, because it seems not able to apply to latent space? If not, how to apply this grid pooling on latent space? I didn't get it from reading paper.

Gofinge commented 3 months ago

It should be available to any point cloud at an encoding stage. Check our PTv2 & PTv3 code. https://github.com/Pointcept/Pointcept/blob/main/pointcept/models/point_transformer_v2/point_transformer_v2m2_base.py#L229

Litoda commented 3 months ago

Does this mean that during grid pooling at any stage, the coordinates of the initial points will follow along with the features?

Gofinge commented 3 months ago

Yes, coordinates go with mean pooling.

Litoda commented 3 months ago

Thank you so much!!!