VoroMesh: Learning Watertight Surface Meshes with Voronoi Diagrams

Nissim Maruani, Roman Klokov, Maks Ovsjanikov, Pierre Alliez, Mathieu Desbrun.

PyTorch implementation of the ICCV2023 paper (project page)

Requirements

The code is tested on the listed versions but other versions may also work:

• Python 3.9
• PyTorch 1.12.1
• PyTorch3D 0.7.0
• MinkowskiEngine 0.5.4
• trimesh 3.15.5
• igl 2.2.1
• tqdm 4.64.1
• yaml 0.2.5
• CGAL (Optional, for fast and accurate VoroMesh extraction)

Compiling C++ utilities (optional)

In theory, our VoroMesh is guaranteed to be watertight. In practice, however, discretization can potentialy create self-intersections. To alleviate this problem, we use CGAL EPIC (exact predicate inexact construction) kernel for both our VoroMesh construction and self-intersection check. In our tests, 100% of our meshes are watertight. These tools can be compiled in src/cpp_utils.

If the C++ utilities are not compiled, the code will still run but the output shapes will probably NOT BE WATERTIGHT.

Direct Optimization

Our optimized models can be downloaded here.

To fit a VoroMesh to a single shape:

python src/single_voromesh.py ./data/bunny.stl

To process the Thingi32 dataset with resolutions 32, 64 and 128:

python src/batch_voromesh.py

To compute the relevant metrics:

python src/eval/eval_THINGI.py out/direct_voromesh_32/
python src/eval/eval_THINGI.py out/direct_voromesh_64/
python src/eval/eval_THINGI.py out/direct_voromesh_128/

Datasets and pre-trained weights

Preprocessed datasets ABC (10.4 Gb) and Thingi32 (0.7 Gb) should be unpacked in the ./data directory, pretrained model checkpoints (0.4 Gb) - in the ./ directory of the project.

Learning-based reconstruction from SDF input

To evaluate our pretrained models on ABC with different resolutions:

./scripts/evaluate_voromesh_32_32_ABC.sh
./scripts/evaluate_voromesh_64_64_ABC.sh
./scripts/evaluate_voromesh_32+64_32_ABC.sh
./scripts/evaluate_voromesh_32+64_64_ABC.sh

To compute the relevant metrics:

python src/eval/eval_ABC.py samples/ABC_32_32
python src/eval/eval_ABC.py samples/ABC_64_64
python src/eval/eval_ABC.py samples/ABC_32+64_32
python src/eval/eval_ABC.py samples/ABC_32+64_64

To evaluate our pretrained models on Thingi32 with different resolutions:

./scripts/evaluate_voromesh_32_32_thingi32.sh
./scripts/evaluate_voromesh_32_64_thingi32.sh
./scripts/evaluate_voromesh_32_128_thingi32.sh
./scripts/evaluate_voromesh_32+64_32_thingi32.sh
./scripts/evaluate_voromesh_32+64_64_thingi32.sh
./scripts/evaluate_voromesh_32+64_128_thingi32.sh

To compute the relevant metrics:

python src/eval/eval_THINGI.py samples/Thingi32_32 -formatted 1
python src/eval/eval_THINGI.py samples/Thingi32_64 -formatted 1
python src/eval/eval_THINGI.py samples/Thingi32_128 -formatted 1
python src/eval/eval_THINGI.py samples/Thingi32_32+64_32 -formatted 1
python src/eval/eval_THINGI.py samples/Thingi32_32+64_64 -formatted 1
python src/eval/eval_THINGI.py samples/Thingi32_32+64_128 -formatted 1

To train with different input SDF resolutions:

./scripts/train_voromesh_32.sh
./scripts/train_voromesh_64.sh
./scripts/train_voromesh_32+64.sh

Watertight check (requires C++ utilities)

To check for watertightness:

python src/eval/check_watertight.py out/direct_voromesh_32

Citation

If you find our work useful in your research, please consider citing:

@InProceedings{maruani23iccv,
  author    = {N. Maruani and R. Klokov and M. Ovsjanikov and P. Alliez and M. Desbrun},
  title     = {VoroMesh: Learning Watertight Surface Meshes with Voronoi Diagrams},
  booktitle = {ICCV},
  year      = {2023},
}