search

cszyzhang / RISurConv

Official codes for ECCV2024 paper: RISurConv: Rotation Invariant Surface Attention-Augmented Convolutions for 3D Point Cloud Classification and Segmentation
MIT License
16 stars 1 forks source link

Rotation #4

Open SelenaClover22 opened 2 weeks ago

SelenaClover22 commented 2 weeks ago

Thanks for the great work!

I have one question when trying to replicate your results: in the paper, it shows that the approach demonstrates great stability for rotations, but when I'm checking the code(e.g. for ModelNet), there does not seem to be rotation procedures on the training or test data, and the preprocessing is just for sampling and normalization. I did see some rotation functions in provider.py, but I did not find where you used them to train the e.g. classification task.

SelenaClover22 commented 2 weeks ago

and one more issue: PyTorch 1.9.0 is not compatible with CUDA 12.1, I wonder if another version of PyTorch or CUDA was used..