Skip to content

[SGP 2021] Scalable Surface Reconstruction with Delaunay-Graph Neural Networks

License

Notifications You must be signed in to change notification settings

raphaelsulzer/dgnn

Repository files navigation

Scalable Surface Reconstruction with Delaunay-Graph Neural Networks [SGP 2021]

MVS point cloud Ours ConvONet [1]

Paper | Video

This repository contains the implementation of the paper:

Scalable Surface Reconstruction with Delaunay-Graph Neural Networks
Raphael Sulzer, Loic Landrieu, Renaud Marlet, Bruno Vallet
SGP 2021

If you find our code or paper useful, please consider citing

@article{sulzer2021dgnn,
   title={Scalable Surface Reconstruction with Delaunay‐Graph Neural Networks},
   volume={40},
   ISSN={1467-8659},
   url={http://dx.doi.org/10.1111/cgf.14364},
   DOI={10.1111/cgf.14364},
   number={5},
   journal={Computer Graphics Forum},
   publisher={Wiley},
   author={Sulzer, R. and Landrieu, L. and Marlet, R. and Vallet, B.},
   year={2021},
   month={Aug},
   pages={157–167}
}

Installation

Please follow the instructions step-by-step.

  1. Clone the repository to your local machine and enter the folder
git clone git@github.com:raphaelsulzer/dgnn.git
  1. Create an anaconda environment called dgnn
conda env create -f environment.yaml
conda activate dgnn
  1. Compile the extension module libmesh (taken from Convolutional Occupancy Networks [1])
cd utils
python setup_libmesh_convonet.py build_ext --inplace

Surface mesh reconstruction using our pretrained models

Berger et al. dataset

Reconstruct the Berger et al. [2] dataset from the scans used in our paper.

  1. Download and unzip the dataset in the data folder
bash data/download_reconbench.sh
  1. Reconstruct and evaluate the meshes
python run.py -i --config configs/pretrained/reconbench.yaml

ModelNet10

Reconstruct and evaluate ModelNet10 from Deep Surface Reconstruction from Point Clouds with Visibility Information

  1. Download and unzip the dataset in the data folder
bash data/download_modelnet.sh
  1. Reconstruct and evaluate the meshes
python run.py -i --config configs/pretrained/modelnet.yaml

Custom object or scene

To reconstruct any object or scene from a point cloud you need a pointcloud.npz file. See Custom dataset for the structure of the file. Additionally you need to adjust the configs/custom.yaml file to point to your dataset.

Once you have everything set up run

python run.py -i --config configs/custom.yaml

Training

Training a new model requires closed ground truth meshes. You can use the procedure explained in Custom dataset to construct a new dataset from such meshes. Once you have created the dataset you need to add your dataset to the getDataset() function in processing/dataset.py and prepare a custom.yaml file (see configs/reconbench.yaml for an example). The .yaml file also allows to change several model parameters.

To train a new model run

python run.py -t --config configs/custom.yaml

Custom dataset

To train or reconstruct a custom dataset you need to provide for each object or scene a pointcloud.npz file containing a points, normals and sensor_position field (see data/reconbench/3_scan/ for examples of such files). The normals field is not actually used in this work but has to be present in the file. It can e.g. be set to contain zero vectors. The datatype of all fields has to be np.float64.

Furthermore you need several files containing the 3D Delaunay triangulation and feature information for each object or scene.

We provide prebuild binaries for Ubuntu 18.04 to synthetically scan a mesh, build a 3D Delaunay triangulation, and extract ground truth labels and features. Note that this uses the scanning procedure introduced in DSRV and not the one from Berger et al. [2] used in our paper.

Scan ground truth mesh

utils/scan -w path/to/workingDir -i filenameMeshToScan --export npz

Extract labels and features

utils/feat -w path/to/workingDir -i filenameScan -o filenameOutput -g filenameGroundTruthMesh -s npz

See processing/reconbench/feat.py and processing/reconbench/scan.py for examples on how to batch process all files in your dataset.

If you want further information about the tools used for scanning and feature extraction or you want to compile them youself have a look at the mesh-tools repository.

References

[1] PENG S., NIEMEYER M., MESCHEDER L., POLLEFEYS M., GEIGER A.: Convolutional occupancy networks. In ECCV (2020).

[2] BERGER M., LEVINE J. A., NONATO L. G., TAUBIN G., SILVA C. T.: A benchmark for surface reconstruction. ACM Transaction on Graphics. (2013).

[3] SCHÖPS T., SCHÖNBERGER J. L., GALLIANI S., SATTLER T., SCHINDLER K., POLLEFEYS M., GEIGER A.: A multi-view stereo benchmark with high-resolution images and multi-camera videos. In CVPR (2017).