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
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}
}
ModelNet40
and ScanObjectNN
are available now. model.py
by adding class ABEM_Module(nn.Module)
.Train the model:
modelnet40_ply_hdf5_2048
folder to ./data
main.py
):
python main.py --exp_name=gbnet_modelnet40 --model=gbnet --dataset=modelnet40
Test the pre-trained model:
./pretrained
python main.py --exp_name=gbnet_modelnet40_eval --model=gbnet --dataset=modelnet40 --eval=True --model_path=pretrained/gbnet_modelnet40.t7
Train the model:
training_objectdataset_augmentedrot_scale75.h5
and test_objectdataset_augmentedrot_scale75.h5
files to ./data
main.py
):
python main.py --exp_name=gbnet_scanobjectnn --model=gbnet --dataset=ScanObjectNN
Test the pre-trained model:
./pretrained
python main.py --exp_name=gbnet_scanobjectnn_eval --model=gbnet --dataset=ScanObjectNN --eval=True --model_path=pretrained/gbnet_scanobjectnn.t7
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
The code is built on DGCNN. We thank the authors for sharing the codes.