search

filaPro / oneformer3d

[CVPR2024] OneFormer3D: One Transformer for Unified Point Cloud Segmentation
Other
348 stars 32 forks source link

Implementation on windows #20

Closed BaptFontaine closed 7 months ago

BaptFontaine commented 10 months ago

I tried running your test script on the s3dis dataset but ran into trouble installing MinkowskiEngine on Windows. When I attempted to install it, I got a message saying it's not supported on Windows. This happened in the oneformer3d.py file. I'm on Windows 10 with Torch 1.13 and CUDA 11.6. Could you provide guidance on how to either install MinkowskiEngine on Windows or suggest an alternative solution that doesn't require it?

Thank you for your really interesting work

filaPro commented 10 months ago

I can only recommend to search windows in MinkowskiEngine github repo. Looks like there are some solutions.

Also the backbone for S3DIS dataset is based SpConv, so we don't need much MinkowskiEngine. The only thing is S3DISOneFormer3D.collate function. We use MinkowskiEngine there for voxelizing the input point cloud. It should be possible to do this step with SpConv.

BaptFontaine commented 9 months ago

Thank you for your response. I will try to figure it out, but it seems that the Minkowski Engine isn't suitable for Windows.

I have another question regarding continuing to build my implementation. Your configuration file from S3DISOneformer refers to a SpConvUNet backbone model, but I can't figure out where to find it. Could you please tell me where it comes from?

filaPro commented 9 months ago

I think here.

BaptFontaine commented 9 months ago

Hi, I attempted to utilize SpConv for the S3DISOneFormer3D.collate function, but I am unsure how to correctly replace lines 323 and 324. Is there an easy solution to replace these lines with SpConv?

filaPro commented 9 months ago

One way is probably to use SPFomer code as we adapt exactly the same pre-trained backbone as they for ScanNet. So their pre-processing without MinkowskiEngine is exactly the same as ours.