PointNorm-for-Point-Cloud-Analysis

This is the official Pytorch implementation of our paper "PointNorm: Dual Normalization is All You Need for Point Cloud Analysis"

Updates

• 2023/4/7: Our paper is accepted by IJCNN 2023 (Oral)!

• 2022/9/15: We have revised the paper and have updated the manuscript for arxiv. Here is a non-exhausted list of the change.

  • Change the format from CVF to IEEE style
  • Add the experiments for S3DIS semantic segmentation
  • Add Optimization Landscape Analysis
  • Merge the figure for shape classification and semantic segmentation.
  • Reformat the mathematical notations and deviations.
  • Update the instruction on Readme.

• 2022/8/5: Code for visualizing the loss landscape and the segmentation output has been uploaded. Detailed instructions will be included in this readme file next week.

• 2022/7/28: Code for classifcation and segmentation have been uploaded.

• 2022/7/24: Basic instructions have been given. The complete code will be uploaded next week.

• 2022/7/15: Arxiv Link is available at: https://arxiv.org/abs/2207.06324

• 2022/7/13: Stay tuned. Code will be uploaded soon.

Abstract

Point cloud analysis is challenging due to the irregularity of the point cloud data structure. Existing works typically employ the ad-hoc sampling-grouping operation of PointNet++, followed by sophisticated local and/or global feature extractors for leveraging the 3D geometry of the point cloud. Unfortunately, the sampling-grouping operations do not address the point cloud's irregularity, whereas the intricate local and/or global feature extractors led to poor computational efficiency. In this paper, we introduce a novel DualNorm module after the sampling-grouping operation to effectively and efficiently address the irregularity issue. The DualNorm module consists of Point Normalization, which normalizes the grouped points to the sampled points, and Reverse Point Normalization, which normalizes the sampled points to the grouped points. The proposed framework, PointNorm, utilizes local mean and global standard deviation to benefit from both local and global features while maintaining a faithful inference speed. Experiments show that we achieved excellent accuracy and efficiency on ModelNet40 classification, ScanObjectNN classification, ShapeNetPart Part Segmentation, and S3DIS Semantic Segmentation.

Intuition

Workflow

Standard Deviation Analysis

Dependencies

• Python 3.8
• Pytorch 1.9
• CUDA 11
• torchvision 0.10

Install

• Install the required packages.

  # clone this repository
  git clone https://github.com/ShenZheng2000/PointNorm-for-Point-Cloud-Analysis.git
  cd PointNorm-for-Point-Cloud-Analysis
  
  # install required packages
  pip install cycler einops h5py pyyaml==5.4.1 scikit-learn==0.24.2 scipy tqdm matplotlib==3.4.2
  pip install pointnet2_ops_lib/.
  

Classification on ModelNet40

• The results at ModelNet40 are volatile. Running the same model with different seeds leads to significantly different scores. To alleviate the randomness, you may consider training the model with different seeds for 2-4 times and report the average accuracy as your score.

• Different methods use different voting strategies to promote their classification accuracy (see the paper for details). For fairness, we do not use any voting strategy in our experiments.

• Train on ModelNet40.

  cd classification_ModelNet40
  
  # For PointNorm (the full-sized model), run:
  python main.py --model PointNorm_2_2
  
  # For PointNorm-Tiny (the lightweight model), run:
  python main.py --model PointNormTiny --embed_dim 32 --res_expansion 0.25
  

Classification on ScanObjectNN

• Train on ScanObjectNN.

  cd classification_ScanObjectNN
  
  # For PointNorm (the full-sized model), run:
  python main.py --model PointNorm_2_2 --point_norm True --reverse_point_norm True --local_mean True --global_std True
  
  # For PointNorm-Tiny (the lightweight model), run:
  python main.py --model PointNormTiny --point_norm True --reverse_point_norm True --local_mean True --global_std True --embed_dim 32 --res_expansion 0.25
  

• Calculate the loss landscape for model trained on ScanObjectNN to generate loss_landscape.txt.

  python lossland.py --model PointNorm --msg {the_checkpoint_message} --point_norm True --reverse_point_norm True --local_mean True --global_std True
  

• Add the path for all loss_landscape.txt in filelist in plot_landscape.py

• Plot the loss landscape.

  python plot_landscape.py
  

Part segmentation on ShapeNetPart

• Prepare ShapeNetPart Dataset.

  cd part_segmentation
  mkdir data
  cd data
  wget https://shapenet.cs.stanford.edu/media/shapenetcore_partanno_segmentation_benchmark_v0_normal.zip --no-check-certificate
  unzip shapenetcore_partanno_segmentation_benchmark_v0_normal.zip
  

• Train on ShapeNetPart.

  cd part_segmentation
  
  # For PointNorm (the full-size model), run:
  python main.py --model PointNorm --point_norm True --reverse_point_norm True --local_mean True --global_std True
  
  # For PointNorm-Tiny (the lightweight model), run:
  python main.py --model PointNormTiny --point_norm True --reverse_point_norm True --local_mean True --global_std True --embed_dim 32 --res_expansion 0.25
  

• Visualize the part segmentation result on ShapeNetPart.

  # For result visualization, run:
  python plot_ptseg.py --model {PointNorm/PointNormTiny} --exp_name {your_checkpoint_name} --angle_one {your_elevation_angle} --angle_two {your_azimuth_angle} --id {the_object_id}
  

• Some examples.

  python plot_ptseg.py --model PointNorm --exp_name PointNorm_7_8_v1 --angle_one 330 --angle_two 90 --id 1 # airplane
  python plot_ptseg.py --model PointNorm --exp_name PointNorm_7_8_v1 -angle_one 270 --angle_two 90 --id 500 # car
  python plot_ptseg.py --model PointNorm --exp_name PointNorm_7_8_v1 --angle_one 120 --angle_two 270 --id 1000 # chair
  python plot_ptseg.py --model PointNorm --exp_name PointNorm_7_8_v1 --angle_one 90 --angle_two 270 --id 1500 # lamp
  python plot_ptseg.py --model PointNorm --exp_name PointNorm_7_8_v1 --angle_one 90 --angle_two 270 --id 2000 # skateboard
  

Semantic Segmentation on S3DIS

• Due to a limited amount of GPUs, we only examine PointNet++ (w/ DualNorm) on semantic segmentation tasks. If we have more GPUs in the future, we will try PointNorm.

• Clone the following repository

  git clone https://github.com/yanx27/Pointnet_Pointnet2_pytorch.git
  

• Do the following modifications.

  • Replace train_semseg.py with train_semseg.py
  • Replace test_semseg.py with test_semseg.py
  • Replace pointnet2_sem_seg.py with pointnet2_sem_seg.py
  • Replace pointnet2_utils.py with pointnet2_utils.py
  • Add data_augmentation.py

• Download S3DIS and save in ./data/s3dis/Stanford3dDataset_v1.2_Aligned_Version/

• Process S3DIS (this might takes 10-20 minutes, so be patient.)

  cd data_utils
  python collect_indoor3d_data.py
  

• Find you data folder in ./data/stanford_indoor3d/

• Go to train_semseg.py and test_semseg.py, change root to the path of the data folder.

• Train on S3DIS

  python train_semseg.py --model pointnet2_sem_seg --log_dir {name_for_your_log_dir}
  

• Test on S3DIS

  python test_semseg.py --log_dir {name_for_your_log_dir} --visual # with visualization
  

• Double check the.obj files in ./log/sem_seg/{name_for_your_log_dir}/visual/.

• Visualize the semantic segmentation output using MeshLab.

Optimization Landscape Analysis

• Train PointNorm variants (w or w/o DualNorm) with different learning rates.

  cd classification_ScanObjectNN
  
  # With DualNorm (lr = 0.01, 0.02, 0.005)
  python main_draw.py --learning_rate 0.01 --model PointNorm_2_2 --point_norm True --reverse_point_norm True --local_mean True --global_std True
  python main_draw.py --learning_rate 0.02 --model PointNorm_2_2 --point_norm True --reverse_point_norm True --local_mean True --global_std True
  python main_draw.py --learning_rate 0.005 --model PointNorm_2_2 --point_norm True --reverse_point_norm True --local_mean True --global_std True
  
  # Without DualNorm (lr = 0.01, 0.02, 0.005)
  python main_draw.py --learning_rate 0.01 --model PointNorm_2_2 --local_mean True --global_std True
  python main_draw.py --learning_rate 0.02 --model PointNorm_2_2 --local_mean True --global_std True
  python main_draw.py --learning_rate 0.005 --model PointNorm_2_2 --local_mean True --global_std True  
  

• Record the corresponding log.out in vis_files/plot_comparison.py.

• Visualize the result.

  cd ..
  cd vis_files
  python plot_comparison.py
  

TODO-List

• Upload readme with basic instructions
• Upload files for classfication at ScanObjectNN
• Upload files for segmentation at ShapeNetPart
• Upload files for classfication at ModelNet40
• Update the code for semantic segmentation
• Upload files and instruction for visualization
• Update the code for loss landscape visualization.
• Update the code for Optimization Landscape Analysis.
• Update readme

BibTeX

Please cite our paper if you find this repository helpful.

@article{zheng2022pointnorm,
  title={PointNorm: Dual Normalization is All You Need for Point Cloud Analysis},
  author={Zheng, Shen and Pan, Jinqian and Lu, Changjie and Gupta, Gaurav},
  journal={arXiv preprint arXiv:2207.06324},
  year={2022}
}

Contact

shenzhen@andrew.cmu.edu

Acknowledgment

This repository is heavily based upon PointMLP and PointNet++ . We appreciate their excellent work and kind sharing.