Awesome

Voxel-based Network for Shape Completion by Leveraging Edge Generation

This is the PyTorch implementation for the paper "Voxel-based Network for Shape Completion by Leveraging Edge Generation (ICCV 2021, oral)"

Getting Started

python version: python-3.6; cuda version: cuda-10; PyTorch version: 1.5

Compile Customized Operators

Build operators under ops by using python setup.py install.

Datasets

Our dataset PCN's dataset TopNet's dataset

Train the model

To train the models on pcn dataset: python train_edge.py
--train_pcn;
--loss_type: pcn;
--train_path: the training data;
--eval_path: the validation data;
--n_gt_points: 16384;
--n_out_points: 16384;
--density_weight:1e11;
--dense_cls_weight:1000;
--p_norm_weight:0;
--dist_regularize_weight:0;
--chamfer_weight:1e6;
--lr 0.0007.

To train the models on topnet dataset: python train_edge.py
--train_pcn;
--loss_type: topnet;
--train_path: the training data;
--eval_path: the validation data;
--n_gt_points: 2048;
--n_out_points: 2048;
--density_weight:1e10;
--dense_cls_weight:100;
--p_norm_weight:300;
--dist_regularize_weight:0.3;
--chamfer_weight:1e4;
--augment;
--lr 0.0007.

To train the models on our dataset: python train_edge.py
--train_seen;
--loss_type: topnet;
--h5_train: the training data;
--h5_val: the validation data;
--n_gt_points: 2048;
--n_out_points: 2048;
--density_weight:1e10;
--dense_cls_weight:100;
--p_norm_weight:300;
--dist_regularize_weight:0.3;
--chamfer_weight:1e4;
--lr 0.0007.

Evaluate the models

The pre-trained models can be downloaded here: Models, unzip and put them in the root directory.
To evaluate models: python test_edge.py
--loss_type: topnet or pcn;
--eval_path: the test data from different cases;
--checkpoint: the pre-trained models;
--num_gt_points: the resolution of ground truth point clouds.

Generate edge points

edge_detection.py is used to generate edge points given a point cloud data. In order to run it, feed the data path and you will get the edge points saved in a h5py format.

Citation

@inproceedings{wang2021voxel,
     author = {Wang, Xiaogang and , Marcelo H. Ang Jr. and Lee, Gim Hee},
     title = {Voxel-based Network for Shape Completion by Leveraging Edge Generation},
     booktitle = {ICCV)},
     year = {2021},
}

Acknowledgements

Our implementations use the code from the following repository:
Chamferdistance
PointNet++
convolutional_point_cloud_decoder