Geometric Back-projection Network for Point Cloud Classification

This repository is for Geometric Back-projection Network (GBNet) introduced in the following paper:
Geometric Back-projection Network for Point Cloud Classification
Shi Qiu, Saeed Anwar, Nick Barnes
IEEE Transactions on Multimedia (TMM), 2021

Paper and Citation

The paper can be downloaded from arXiv and IEEE.
If you find our paper/code is useful, please cite:

    @article{qiu2022geometric,
        title={Geometric Back-projection Network for Point Cloud Classification},
        author={Qiu, Shi and Anwar, Saeed and Barnes, Nick},
        journal={IEEE Transactions on Multimedia},
        year={2022},
        volume={24},
        pages={1943-1955},
        doi={10.1109/TMM.2021.3074240}
    }

Network Architecture

Updates

• 23/04/2021 Codes for both ModelNet40 and ScanObjectNN are available now.
• 27/04/2021 Update model.py by adding class ABEM_Module(nn.Module).
• 29/04/2021 Pre-trained model (OA: 80.50%, mAcc: 77.31%) on ScanObjectNN is available at google drive.
• 30/04/2021 Update a pre-trained model (OA: 80.99%, mAcc: 78.21%) on ScanObjectNN via google drive.
• 03/05/2021 Pre-trained model (OA: 93.80%, mAcc: 91.04%) on ModelNet40 is available at google drive.

Implementation Platforms

• Python 3.6
• Pytorch 0.4.0 with Cuda 9.1
• Higher Python/Pytorch/Cuda versions should also be compatible

ModelNet40 Experiment

Train the model:

• download ModelNet40, unzip and move modelnet40_ply_hdf5_2048 folder to ./data
• then run (more settings can be modified in main.py):

python main.py --exp_name=gbnet_modelnet40 --model=gbnet --dataset=modelnet40

Test the pre-trained model:

• put the pre-trained model under ./pretrained
• then run:

python main.py --exp_name=gbnet_modelnet40_eval --model=gbnet --dataset=modelnet40 --eval=True --model_path=pretrained/gbnet_modelnet40.t7

ScanObjectNN Experiment

Train the model:

• download ScanObjectNN, and extract both training_objectdataset_augmentedrot_scale75.h5 and test_objectdataset_augmentedrot_scale75.h5 files to ./data
• then run (more settings can be modified in main.py):

python main.py --exp_name=gbnet_scanobjectnn --model=gbnet --dataset=ScanObjectNN

Test the pre-trained model:

• put the pre-trained model under ./pretrained
• then run:

python main.py --exp_name=gbnet_scanobjectnn_eval --model=gbnet --dataset=ScanObjectNN --eval=True --model_path=pretrained/gbnet_scanobjectnn.t7

Pre-trained Models

• Python 3.6, Pytorch 0.4.0, Cuda 9.1
• 2 Nvidia P100 GPUs
• using default training settings as in main.py

Model	Dataset	#Points	Data Augmentation	Loss	Performance on Test Set	Download Link
GBNet	ModelNet40	1024	random scaling and translation	cross-entropy with label smoothing	overall accuracy: 93.80% average class accuracy: 91.04%	google drive
GBNet	ScanObjectNN	1024	random scaling and translation	cross-entropy with label smoothing	overall accuracy: 80.99% average class accuracy: 78.21%	google drive

For more discussions regarding the factors that may affect point cloud classification,
please refer to the following paper:
Revisiting Point Cloud Classification with a Simple and Effective Baseline

Acknowledgement

The code is built on DGCNN. We thank the authors for sharing the codes.