Multimodal-Shape-Completion

This repository provides PyTorch implementation of our paper:

Multimodal Shape Completion via Conditional Generative Adversarial Networks

Rundi Wu*, Xuelin Chen*, Yixin Zhuang, Baoquan Chen (* equal contribution)

ECCV 2020 (spotlight)

Prerequisites

• Linux
• NVIDIA GPU + CUDA CuDNN
• Python 3.6

Dependencies

1. Install python package dependencies through pip:

   pip install -r requirements.txt

2. Install external dependency PyTorchEMD for the usage of EMD loss.

Data

We use three datasets in our paper.

1. 3D-EPN

   Please download the partial scan point cloud data from their website and extract it into data folder. For the complete point clouds data, please download it from PKU disk and extract it into data folder. Or you can follow this �blender-render script to virtually scan ShapeNet objects by yourself.

2. PartNet

   Please first download the PartNet dataset and then run our script to merge the point cloud segmentation label to the first level:

   cd data/partnet_process
   python pc_merge.py --src {your-path-to-PartNet-dataset}

   Also, remember to change the data_root path in the corresponding scripts for training/testing on PartNet dataset.

3. PartNet-Scan

   Please first download the PartNet dataset and then run our script to (1) create 4 parital mesh for each complete shape and (2) scan complete shape and parital mesh using blender.

   cd data/partnet_process
   sh gen_partnet_scan.sh {your-path-to-PartNet-dataset} {your-path-to-save-scanned-results}

   Note that the scan process (blender render) may take a very long time and the outputs require quite a lot disk space. You can also increase the number of sub-process (default 10) to speed up.

Training

Training scripts can be found in scripts folder and please see common.py for specific definition of all command line parameters. For example, to train on 3DEPN chair category:

# 1. pre-train the PointNet AE
sh scripts/3depn/chair/train-3depn-chair-ae.sh

# 2. pre-train the PointNet VAE
sh scripts/3depn/chair/train-3depn-chair-vae.sh

# 3. train the conditional GAN for multimodal shape completion
sh scripts/3depn/chair/train-3depn-chair-gan.sh

Training log and model weights will be saved in proj_log folder by default.

Testing

Testing scripts can also be found in scripts folder. For example, to test the model trained on 3DEPN chair category:

# by default, run over all test examples and output 10 completion results for each
sh scripts/3depn/chair/test-3depn-chair-gan.sh

The completion results, along with the input parital shape, will be saved in proj_log/mpc-3depn-chair/gan/results by default.

Evaluation

Evaluation code can be found in evaluation folder. To evaluate the completion diversity , fidelity and quality:

cd evaluation/
# calculate Total Mutual Difference (TMD)
python total_mutual_diff.py --src {path-to-saved-testing-results}
# calculate Unidirectional Hausdorff Distance (UHD) and completeness
python completeness.py --src {path-to-saved-testing-results}
# calculate Minimal Matching Distance (MMD), this requries a tensorflow environment
python mmd.py --src {path-to-saved-testing-results} --dataset {which-dataset} --class_name {which-category} -g 0

Note that MMD calculation requires an compiled external library from its original repo.

Pre-trained models

Pretrained model can be downloaded from PKU disk or Google Drive. Put the downloaded files into proj_log folder and then testing scripts shall be run directly.

Cite

Please cite our work if you find it useful:

@InProceedings{wu_2020_ECCV,
author = {Wu, Rundi and Chen, Xuelin and Zhuang, Yixin and Chen, Baoquan},
title = {Multimodal Shape Completion via Conditional Generative Adversarial Networks},
booktitle = {The European Conference on Computer Vision (ECCV)},
month = {August},
year = {2020}
}