Our code is located in DSConv-master/openpoints/models/backbone/, where pointnext.py and pointconvnext.py represent our algorithms. pointnext.py has not modified the module name and can directly load the .pth files we provided. pointconvnext.py is our official version of the code, where we have modified the module names according to the paper. It includes comments throughout the algorithm's flow, primarily in LocalAggregation. However, it cannot directly load the .pth files.

The provided .pth files contain unused parameters during training (such as LocalAggregation.convs, which can be commented out without affecting test results). Due to the presence of these parameters, directly using pointnext.py for parameter counting may yield results inconsistent with the paper. Therefore, please use pointconvnext.py to test the number of parameters, GFLOPs, and throughput.

To switch between these two code versions, please use different config files in the cfg directory. We default to providing code for segmentation tasks, but if you need to test classification tasks, please replace the existing files with the two Python files in backbone/classification.

For specific instructions on using the code, please refer to the official pointnext codebase for environment configuration and compilation. During runtime, use the cfg directory we provided and overwrite the "models" and "optim" sections of the official openpoints with our openpoints.

Use_amp is not applicable in this code, please do not enable it.

We have renamed our method from PointConvNext to DSConv, but due to time constraints, we have not fully modified the code. For a better experience, you can directly review DSConv-XL.
