Awesome

Cascaded Refinement Network for Point Cloud Completion

This is the TensorFlow implementation for the paper "Cascaded Refinement Network for Point Cloud Completion"

Getting Started

python version: python-3.5; cuda version: cuda-10; TensorFlow version: tensorflow-1.13

Compile Customized TF Operators

Please follow PointNet++ to compile customized TF operators. You may need to change the related library path if necessary.

Datasets

Our dataset PCN's dataset TopNet's dataset

Train the model

To train the models: python train.py
--loss_type: CD_T or CD_P;
--h5_train: the training data;
--h5_val: the validation data;
--num_gt_points: the resolution of ground truth point clouds;
--step_ratio: 2, 4, 8, 16 for different output resolutions;
--log_dir: output directory to save models and training log;
--augment: whether to use data augmentation during training.

Evaluate the models

Our pre-trained models can be downloaded here: Models, unzip and put them in the root directory.
To evaluate models in different cases: python test.py
--loss_type: CD_T or CD_P;
--data_dir: the test data from different cases;
--checkpoint: the pre-trained models;
--num_gt_points: the resolution of ground truth point clouds;
--step_ratio: 2, 4, 8, 16 for different output resolutions.

Citation

@inproceedings{Wang_2020_CVPR,
     author = {Wang, Xiaogang and , Marcelo H. Ang Jr. and Lee, Gim Hee},
     title = {Cascaded Refinement Network for Point Cloud Completion},
     booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
     month = {June},
     year = {2020},
}

Acknowledgements

Our implementations use the code from the following repository:
PCN
PointNet++