search

LiyaoTang / contrastBoundary

Contrastive Boundary Learning for Point Cloud Segmentation (CVPR2022)
MIT License
139 stars 11 forks source link

RandLA-Net implementation details #31

Open mariusud opened 1 year ago

mariusud commented 1 year ago

Hi and thank you for sharing you code. I see that you have not yet included code for your RandLa-Net implementation, and have a question regarding to how you implemented this, as it is not deeply explained in your paper.

  1. Did you experiment with using the k-nearest neighbors created in the pre-processing step or did you simply query using radius=0.01 as you did you ConvNet?
  2. Did you implement a maximum number of points in the defined local Neighborhood N_i for each point?
LiyaoTang commented 8 months ago

Hi,

Thanks for your interest and sorry for my late response.

For RandLA-Net, we follow its implementation here. That is,

  1. we query with kNN
  2. for the local neighborhood, we also use kNN with k=16 (which I believe is the same as the official implementation but I did not check)

Besides, we recently released another work at nips'23 here that includes our re-implementation of RandLA-Net and should be compatible with this repo. But I still encourage you to check against the original RandLA-Net implementation.

Thanks, Liyao