search

dvlab-research / Stratified-Transformer

Stratified Transformer for 3D Point Cloud Segmentation (CVPR 2022)
MIT License
376 stars 40 forks source link

K-sampling strategies #76

Closed OPradelle closed 1 year ago

OPradelle commented 1 year ago

Hi, thanks for sharing your work.

I have a question about the K-sampling strategies for the stratified version. From what I understand, you compute attention weight for the query over the points set from the dense cloud with smaller windows and the low-density cloud with bigger windows.

However, I don't understand how do you apply those weights to the V vector. Does the V vector contain the encoded value for the denser and low density point set?

Thanks for your answer.

X-Lai commented 1 year ago

Yes, you are correct. The V vector contains the same number of tokens as the keys K, which is composed of both the denser and the sparser point set.