Learning Delaunay Surface Elements for Mesh Reconstruction (DSE meshing)

This is our implementation of the paper "Learning Delaunay Surface Elements for Mesh Reconstruction" at CVPR 2021 (oral), a method for mesh recontruction from a point cloud.

This code was written by Marie-Julie Rakotosaona.

The triangle selection code is based on on Meshing-Point-Clouds-with-IER (with a few smaller modifications) that also uses code from the project annoy. Our network code is based on PointNet.

Prerequisites

CUDA and CuDNN (changing the code to run on CPU should require few changes)
Python 3.6
Tensorflow 1.15

Setup

Install required python packages, if they are not already installed:

pip install numpy
pip install scipy
pip install trimesh

Clone this repository:

git clone https://github.com/mrakotosaon/dse-meshing.git
cd dse-meshing

Setup the triangle selection step:

git submodule update --init --recursive
cd triangle_selection/postprocess
mkdir build
cd build
cmake ..
make

Data

Training set: We store our training data in .tfrecords files. The files contain elements of size N_NEIGHBORSx5 where the first 3 columns contain the 3D coordinates and the last two columns contain the ground truth logmap 2D coordinates.
Pre-trained models on the famousthingi dataset.
Testset from famousthingi dataset.

Our data can be downloaded directly here:

training data: https://nuage.lix.polytechnique.fr/index.php/s/gmnGHjNq7WKipRA
pretrained models: https://nuage.lix.polytechnique.fr/index.php/s/FTCyp5WHg7Z68EM
testing data: https://nuage.lix.polytechnique.fr/index.php/s/3ZcFtqKm6Z27ZJ6

To download our data from the code:

Download pretrained models:

cd data
python download_data.py --task models

Download training set:

cd data
python download_data.py --task training

Download testing set:

cd data
python download_data.py --task testing

Training

To train the classifier network on the provided dataset:

cd train_logmap
python train_classifier.py

To train the logmap estimation network on the provided dataset:

cd train_logmap
python train_logmap_network.py

The trained models are saved in log/log_famousthingi_classifier and log/log_famousthingi_logmap by default. Paths for the training set, output and training parameters can be changed directly in the code.
Training curves are generated during training and can be viewed using tensorboard.

Creating a new training set

To create a new training set please refer to the directory training_set_processing.

Testing

Our testing pipeline has 3 main steps:

Logmap estimation: we locally predict the logmap that contains neighboring points at each point using the trained networks.
Logmap alignment: we locally align the log maps to one another to ensure better consistency.
Triangle selection: we select the produced triangles to generate an almost manifold mesh.

We provide one example point cloud in data/test_data. For easily evaluating your pointclouds or the pointclouds from the testset, move the .xyz files there.

To run all steps on the point clouds in data/test_data directory:

 ./run.sh

The produced meshes can be found in data/test_data/select in the format: final_mesh_(SHAPE_NAME).ply. For numerical precision reasons we suggest to use point clouds with a bounding box diagonal larger than 1. By default our code evaluates the .xyz point clouds in data/test_data please adapt the paths in the code if you wish to evaluate point clouds at different locations.

1. Logmap estimation

To run only the logmap estimation networks on the point clouds in data/test_data directory:

cd logmap_estimation
python eval_network.py

2. Logmap alignment

To only align the logmap patches from step 1 and compute the appearance frequency for step 3:

cd logmap_alignment
python align_patches.py
python eval_align_meshes.py

3. Triangle selection

To apply triangle selection on the output from step 2:

cd triangle_selection
python select.py

Citation

If you use our work, please cite our paper.

@InProceedings{Rakotosaona_2021_CVPR,
    author    = {Rakotosaona, Marie-Julie and Guerrero, Paul and Aigerman, Noam and Mitra, Niloy J. and Ovsjanikov, Maks},
    title     = {Learning Delaunay Surface Elements for Mesh Reconstruction},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {22-31}
}

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. For any commercial uses or derivatives, please contact us.

maphysart / dse-meshing