Rotation Transformation Network: Learning View-Invariant Point Cloud for Classification and Segmentation
We propose a rotation transformation network for point cloud analysis, called RTN, which could reduce the RDF of input 3D objects to 0. The RTN could be seamlessly inserted into many existing DNNs for point cloud analysis.
[PDF]
This repository is the author's re-implementation of RTN. Extensive experimental results on 3D point cloud classification and segmentation tasks demonstrate that the proposed RTN could improve the performances of several state-of-the-art methods significantly.
Further information please contact Shuang Deng and Qiulei Dong.
Please cite this paper if you want to use it in your work:
@inproceedings{2021rtn,
title={Rotation Transformation Network: Learning View-Invariant Point Cloud For Classification And Segmentation},
author={Deng, Shuang and Liu, Bo and Dong, Qiulei and Hu, Zhanyi},
booktitle={Proceedings of the IEEE International Conference on Multimedia and Expo (ICME)},
pages={1-6},
year={2021},
doi={10.1109/ICME51207.2021.9428265}
}
Download the HDF5 files of ModelNet40 (416MB) and put the folder modelnet40_ply_hdf5_2048 to ./align_modelnet and $your_data_folder.
Align the ModelNet40 dataset:
cd align_modelnet
python data_prepare.py
, and put the generated folder aligned_modelnet40_ply_hdf5_2048 to $your_data_folder.
Train the RTN.
python train_learning_rotation_so3.py --gpu $your_gpu_id --data_path $your_data_folder --log_dir $your_RTN_model_path
Train the RTN+DGCNN.
python train_rotation_pretrain_so3.py --gpu $your_gpu_id --data_path $your_data_folder --transformer_model_path $your_RTN_model_path --log_dir $your_model_path
Test the RTN.
python evaluate_learning_rotation_so3.py --gpu $your_gpu_id --data_path $your_data_folder --model_path $your_RTN_model_path --dump_dir $your_RTN_test_path
Test the RTN+DGCNN.
python evaluate_rotation_pretrain_so3.py --gpu $your_gpu_id --data_path $your_data_folder --transformer_model_path $your_RTN_model_path --model_path $your_model_path --dump_dir $your_test_path
The trained models are stored in the folder result.
The structure of this codebase is borrowed from DGCNN.