PV-Ada: Point-Voxel Adaptive Feature Abstraction for Robust Point Cloud Classification

This repository contains PyTorch implementation for PV-Ada: Point-Voxel Adaptive Feature Abstraction for Robust Point Cloud Classification.

News

• 2022-10 PV-Ada (technical report) is released on arXiv.
• 2022-10 A 5-minute presentation is given in ECCV'22 Workshop on Sensing, Understanding and Synthesizing Humans.
• 2022-09 Our solution based on PV-Ada won the second place in ModelNet-C classification track of PointCloud-C Challenge 2022 (ECCV Workshop 2022) .

Dataset

• ModelNet40 [download]

  |- modelnet40_ply_hdf5_2048
      | - ply_data_train*.h5 (#5)
      | - ply_data_test*.h5 (#2)
      | - shape_names.txt
      | - train_files.txt
      | - test_files.txt
      | - ply_data_train*.json (#5)
      | - ply_data_test*.json (#2)
  

• ModelNetC [download]

    |- modelnet_c
        | - clean.h5
        | - add_global*.h5 (#5)
        | - add_local*.h5 (#5)
        | - dropout_global*.h5 (#5)
        | - dropout_local*.h5 (#5)
        | - jitter*.h5 (#5)
        | - rotate*.h5 (#5)
        | - scale*.h5 (#5)
  

• ExtraModelNetC [download]

  |- cls_extra_test_data.h5
  

Train

python train.py --pw --beta 1.0 --modelnet_root your_path_to_modelnet40 --modelnetc_root your_path_to_modelnetc --[tapering]

e.g. 
python train.py --pw --beta 1.0 --modelnet_root /mnt/ssd1/lifa_rdata/cls/modelnet40_ply_hdf5_2048 --modelnetc_root /mnt/ssd1/lifa_rdata/PointCloud-C/modelnet_c

python train.py --pw --beta 1.0 --modelnet_root /mnt/ssd1/lifa_rdata/cls/modelnet40_ply_hdf5_2048 --modelnetc_root /mnt/ssd1/lifa_rdata/PointCloud-C/modelnet_c --tapering

Evaluate

# For ModelNet40 test set
python evaluate.py --eval --ckpt your_path/model.t7 --modelnet_root your_path_to_modelnet40 

e.g.
python evaluate.py --eval --ckpt pretrained/modelnetc.t7 --modelnet_root /mnt/ssd1/lifa_rdata/cls/modelnet40_ply_hdf5_2048


# For ModelNet40-C Public test set
python evaluate.py --eval_corrupt --ckpt your_path/model.t7 --modelnetc_root your_path_to_modelnetc

e.g.
python evaluate.py --eval_corrupt --ckpt pretrained/modelnetc.t7 --modelnetc_root /mnt/ssd1/lifa_rdata/PointCloud-C/modelnet_c

Infer

python test.py --ckpt your_path/model.t7 --h5_path your_path_to_extra_modelnetc.h5 --saved_path your_path

e.g.
python test.py --ckpt pretrained/modelnetc.t7 --h5_path /mnt/ssd1/lifa_rdata/PointCloud-C/cls_extra_test_data.h5 --saved_path results/PCC 

Citation

If you find our work useful in your research, please consider citing:

Acknowledgements

ModelNet40-C, PointCloud-C, PCT and 3DeformRS.