SPE-Net: Boosting Point Cloud Analysis via Rotation Robustness Enhancement

This repository is for SPE-Net: Boosting Point Cloud Analysis via Rotation Robustness Enhancement. The repository is based on the open-source codebase Detectron2 and CloserLook3D.

Requiremenets

• Linux with Python ≥ 3.6
• PyTorch ≥ 1.8
• fvcore
• pycocotools
• Java 1.8.0

Installation

• Prepare the target dataset following the instructions from the codebase CloserLook3D, and put the pre-processed data in the folder ./dataset.
• Run pytorchpoints/init.sh to compile the C++ code.

Train SPE-Net Model

configs/spe_net_modelnet40_so3.yaml is the config file for training SPE-Net model on ModelNet40. Run script python3 train_net.py --num-gpus 4 --config-file configs/spe_net_modelnet40_so3.yaml.

pytorchpoints/modeling/backbones/resnet.py includes the implementation for SPE-Net overall architecture.

pytorchpoints/modeling/local_aggregation/spe_mlp.py includes the implementation for SPE-MLP.

Acknowledgements

Thanks the contribution of Detectron2, CloserLook3D and the PyTorch team.