The environment of RandLANet for STPLS3D is the same with the original RandLANet?

[whuhxb]

Hi @RockyatASU

Is the RandLANet for STPLS3D the same with the original RandLANet?
https://github.com/QingyongHu/RandLA-Net?utm_source=catalyzex.com

Have you tested RandLANet for STPLS3D? Or, do you have other settings? Which cuda version?

Thanks a lot.

[whuhxb]

@RockyatASU
I can't successfully compile the following files sh compile_op.sh. The bugs show up like this.

line 15, in
import nearest_neighbors.lib.python.nearest_neighbors as nearest_neighbors
ModuleNotFoundError: No module named 'nearest_neighbors.lib.python.nearest_neighbors'

maybe the environment version problems.

[meidachen]

Hi @whuhxb ,

We were using the RTX3090 on a windows machine with CUDA 11.0 to test the RandLA. Since we didn't make any changes to RandLA itself, but only modified the data-related parameters if you were able to set up the environment for https://github.com/QingyongHu/RandLA-Net?utm_source=catalyzex.com, the same environment should work using the RandLA in this repo.

Best,

[whuhxb]

Hi @meidachen
For RandLA-Net, you only used the RealWorldData for training the network, not the Synthetic data for training the network, right?

[meidachen]

Hi @meidachen For RandLA-Net, you only used the RealWorldData for training the network, not the Synthetic data for training the network, right?

I used both, just like our paper discussed. Originally (a few months ago), I released all point clouds in .txt format (0.3 point spacing) which works with RandLA, but later on, I released a dense version of the data in .ply file with more labels and didn't have a chance to update the RandLA and SCF-net for that (sorry for the inconvenience). Here are the .txt files that I initially tested with: https://drive.google.com/file/d/17nUphM7LpmR3tT9eX0M1SXfIVjidr0-T/view?usp=sharing

We haven't had a chance to fully document the RandLA and SCF-net yet but only did it for KpConv and transformer. I will make better documentation for RandLA and SCF-net in the future.

[whuhxb]

Hi @meidachen

Thanks so much for the answering. I have one question, the provided .txt file data contain both the RealWorldData and the Synthetic data, right? And, I just need to follow your instruction of the data_prepare_STPLS3D with the downloaded .txt data can use both the RealWorldData and the Synthetic data, right?

[meidachen]

Hi @meidachen

Thanks so much for the answering. I have one question, the provided .txt file data contain both the RealWorldData and the Synthetic data, right? And, I just need to follow your instruction of the data_prepare_STPLS3D with the downloaded .txt data can use both the RealWorldData and the Synthetic data, right?

Yes, it should work for both.

[meidachen]

Hi @meidachen Thanks for you detailed answering. If I train RandLA-Net with the .ply format input, could I refer to the KPConv data_prepare_STPLS3D.py to prepare the ply format data?

Another question, I have compared the data_prepare_STPLS3D.py in both RandLA-Net and KPConv, I found that they are same only different with the input data format. As I know, KPConv adopts a sphere to query the input data, not just sub_cloud and KDTree. https://github.com/HuguesTHOMAS/KPConv-PyTorch/blob/master/datasets/S3DIS.py

I have not tried to run RandLA on the ply files myself, but it may work that way.

For running KpConv, you can easily follow the instruction, and data_prepare_STPLS3D.py is not actually used I believe. The actual data preparation happened in https://github.com/meidachen/STPLS3D/blob/main/KPConv-PyTorch/datasets/STPLS3D.py
And data_prepare_STPLS3D.py was just a script I used along the way when preparing the code.

[whuhxb]

Hi @meidachen

Thanks for your kind reply.
I have found that the actual data preparation is in https://github.com/meidachen/STPLS3D/blob/main/KPConv-PyTorch/datasets/STPLS3D.py.

Best.

[whuhxb]

Hi @meidachen

Another question, how to the density of point clouds with different datasets, for example, STPLS3D and SensatUrban? How to determine the grid size in data preparation? For Point Transformer, how to set the voxel size?
Thanks.

[meidachen]

The grid size is usually determined based on the experiments. The paper of SensatUrban also discussed about different ways to vocalize the point cloud.

[fyj534609]

@RockyatASU I can't successfully compile the following files sh compile_op.sh. The bugs show up like this.

line 15, in import nearest_neighbors.lib.python.nearest_neighbors as nearest_neighbors ModuleNotFoundError: No module named 'nearest_neighbors.lib.python.nearest_neighbors'

maybe the environment version problems.

Hello, have you resolved this issue? I'm facing a similar problem and I'm unable to find a solution. Could you kindly assist me with it? Thank you.