SAUM

Tensorflow code for the paper SAUM: Symmetry-Aware Upsampling Module for Consistent Point Cloud Completion.

Hyeontae Son, Young Min Kim

@InProceedings{Son_2020_ACCV,
    author    = {Son, Hyeontae and Kim, Young Min},
    title     = {SAUM: Symmetry-Aware Upsampling Module for Consistent Point Cloud Completion},
    booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
    month     = {November},
    year      = {2020}
}

Prerequisites

Clone this repository

git clone https://github.com/countywest/SAUM.git

Download & Link datasets

• PCN
• TopNet
• KITTI
• mkdir data && ln -s [path to dataset] data/[dataset name]
  • dataset name: pcn, topnet, kitti

Preprocess TopNet dataset (optional)

Since TopNet dataset does not provide the ground truth for test data, we used the provided validation set for testing and picked 600 samples from the training data to use it as a validation set. Followings are instructions for preparing TopNet dataset same as our experimental setting.

• cd [path to TopNet dataset]
• rm -rf train.list test test.list && mv val test && mv val.list test.list
• copy configs/topnet_dataset/*.list to the data directory.
• make val directory(partial, gt) using val.list
• make new train.list with remaining training data.

You can also download preprocessed topnet dataset here.

Install Dependencies

pip install -r requirements.txt

Build TensorFlow Extensions

Please assign appropriate path to the vars (cuda_inc, cuda_lib, nvcc, tf_inc, tf_inc_pub, tf_lib) in fps/tf_sampling_compile.sh & pc_distance/makefile

• cd fps && ./tf_sampling_compile.sh
• cd pc_distance && make

Usage

To train the SAUM attached models,

python train.py --config_path configs/[decoder_name].yaml --log_dir [log_directory]

To evaluate the result in the test set,

python test.py --checkpoint [log_directory] --results_dir [result_directory]

Any hyperparameters can be controlled in the yaml files.

Pretrained Models

The pretrained models on PCN dataset with decoder PCN and TopNet are available.[here]

Acknowledgements

This code is based on the project PCN. Thanks for their great work.