Neural Mesh Simplification

Implementation of the paper Neural Mesh Simplification paper by Potamias et al. (CVPR 2022) and the updated info shared in supplementary material.

This Python package provides a fast, learnable method for mesh simplification that generates simplified meshes in real-time.

Overview

Neural Mesh Simplification is a novel approach to reduce the resolution of 3D meshes while preserving their appearance. Unlike traditional simplification methods that collapse edges in a greedy iterative manner, this method simplifies a given mesh in one pass using deep learning techniques.

The method consists of three main steps:

1. Sampling a subset of input vertices using a sophisticated extension of random sampling.
2. Training a sparse attention network to propose candidate triangles based on the edge connectivity of sampled vertices.
3. Using a classification network to estimate the probability that a candidate triangle will be included in the final mesh.

Features

• Fast and scalable mesh simplification
• One-pass simplification process
• Preservation of mesh appearance
• Lightweight and differentiable implementation
• Suitable for integration into learnable pipelines

Installation

conda create -n neural-mesh-simplification python=3.12
conda activate neural-mesh-simplification
conda install pip

Depending on whether you are using PyTorch on a CPU or a GPU, you'll have to use the correct binaries for PyTorch and the PyTorch Geometric libraries. You can install them via:

pip install torch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 --index-url https://download.pytorch.org/whl/cpu
pip install torch_cluster==1.6.3 torch_geometric==2.5.3 torch_scatter==2.1.2 torch_sparse==0.6.18 -f https://data.pyg.org/whl/torch-2.4.0+cpu.html

Replace “cpu” with “cu121” or the appropriate CUDA version for your system. If you don't know what is your cuda version, run nvidia-smi

After that you can install the remaining requirements

pip install -r requirements.txt
pip install -e .

Example Usage / Playground

1. Drop your meshes as .obj files to the examples/data folder
2. Run the following command

python examples/example.py

3. Collect the simplified meshes in examples/data. The simplified mesh objects file name will be the ones prefixed with simplified_.

Data Preparation

If you don't have a dataset for training and evaluation, you can use a collection from HuggingFace's 3D Meshes dataset. See https://huggingface.co/datasets/perler/ppsurf for more information.

Run the following script to use the HuggingFace API to download

python scripts/download_test_meshes.py

Data will be downloaded in the data/raw folder at the root of the project. You can use --target-folder to specify a different folder.

Once you have some data, you should preprocess it using the following script:

python scripts/preprocess_data.py

You can use the --data_path argument to specify the path to the dataset. The script will create a data/processed

Training

To train the model on your own dataset with the prepared data:

python ./scripts/train.py

By default, the default training config at config/default.yaml will be used. You can override it with --config /path/to/your/config.yaml.
You can override the following config parameters:

• the checkpoint directory specified in the config file (where the model will be saved) with --checkpoint-dir.
• the data path with --data-path if you have your data in a different location.

If the training was interrupted, you can resume it by specifying the path to the previously created checkpoint directory with --resume.
Use --debug to see DEBUG logging.

Evaluation

To evaluate the model on a test set:

python ./scripts/evaluate.py --eval-data-path /path/to/test/set --checkpoint /path/to/checkpoint.pth

By default, the default training config at config/default.yaml will be used. You can override it with --config /path/to/your/config.yaml.

Inference

To simplify a mesh using the trained model:

python ./scripts/infer.py --input-file /path/to/your/mesh.obj --output-file --model-checkpoint /path/to/checkpoint.pth --device cpu

The default feature dimension for point sampler and face classifier is 128, but can be configured with --hidden-dim <num>.
The default feature dimension for edge predictor is 64, but can be configured with --edge-hidden-dim <num>.
If you have a CUDA-compatible GPU, you can specify --device cuda to use it for inference.

Citation

If you use this code in your research, please cite the original paper:

@InProceedings{Potamias_2022_CVPR,
    author    = {Potamias, Rolandos Alexandros and Ploumpis, Stylianos and Zafeiriou, Stefanos},
    title     = {Neural Mesh Simplification},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {18583-18592}
}

License

This project is licensed under the MIT License - see the LICENSE file for details. m