NMC

PyTorch implementation for paper Neural Marching Cubes, Zhiqin Chen, Hao Zhang.

Paper | Supplementary Material

Citation

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

@article{chen2021nmc,
  title={Neural Marching Cubes},
  author={Zhiqin Chen and Hao Zhang},
  journal={ACM Transactions on Graphics (Special Issue of SIGGRAPH Asia)},
  volume = {40},
  number = {6},
  year={2021}
}

Notice

We have implemented Neural Dual Contouring (NDC). NDC is based on Dual Contouring and thus much easier to implement than NMC. It takes the same input as NMC but produces less triangles and vertices (1/8 of NMC, 1/4 of NMC-lite, ≈MC33) with better triangle quality. It runs faster than NMC because it has significantly less values to predict for each cube (1 bool 3 float for NDC, v.s. 5 bool 51 float for NMC), therefore the network size can be significantly reduced. Yet, it cannot reconstruct some cube cases, and may introduce non-manifold edges.

Requirements

• Python 3 with numpy, h5py, scipy and Cython
• PyTorch 1.8 (other versions may also work)

Build Cython module:

python setup.py build_ext --inplace

Datasets and pre-trained weights

For data preparation, please see data_preprocessing.

We provide the ready-to-use datasets here.

• groundtruth.7z

Backup links:

• groundtruth.7z (pwd: 1234)

We also provide the pre-trained network weights.

• weights.7z

Backup links:

• weights.7z (pwd: 1234)

Note that the weights are divided into six folders:

Training and Testing

Before training, please replace LUT_tess.npz (the Look-Up Table for cube tessellations) in the main directory with the corresponding version of your training target (either NMC or NMC-lite). Both versions of LUT_tess.npz can be found at tessellation.

To train/test NMC with SDF input:

python main.py --train_bool --epoch 400 --data_dir groundtruth/gt_NMC --input_type sdf
python main.py --train_float --epoch 400 --data_dir groundtruth/gt_NMC --input_type sdf
python main.py --test_bool_float --data_dir groundtruth/gt_NMC --input_type sdf

To train/test NMC-lite with SDF input:

python main.py --train_bool --epoch 400 --data_dir groundtruth/gt_simplified --input_type sdf
python main.py --train_float --epoch 400 --data_dir groundtruth/gt_simplified --input_type sdf
python main.py --test_bool_float --data_dir groundtruth/gt_simplified --input_type sdf

To train/test NMC with voxel input:

python main.py --train_bool --epoch 200 --data_dir groundtruth/gt_NMC --input_type voxel
python main.py --train_float --epoch 100 --data_dir groundtruth/gt_NMC --input_type voxel
python main.py --test_bool_float --data_dir groundtruth/gt_NMC --input_type voxel

To train/test NMC-lite with voxel input:

python main.py --train_bool --epoch 200 --data_dir groundtruth/gt_simplified --input_type voxel
python main.py --train_float --epoch 100 --data_dir groundtruth/gt_simplified --input_type voxel
python main.py --test_bool_float --data_dir groundtruth/gt_simplified --input_type voxel

To evaluate Chamfer Distance, Normal Consistency, F-score, Edge Chamfer Distance, Edge F-score, you need to have the ground truth normalized obj files ready in a folder objs. See data_preprocessing for how to prepare the obj files. Then you can run:

python eval_cd_nc_f1_ecd_ef1.py

To count the number of triangles and vertices, run:

python eval_v_t_count.py

If you want to test on your own dataset, please refer to data_preprocessing for how to convert obj files into SDF grids and voxel grids. If your data are not meshes (say your data are already voxel grids), you can modify the code in utils.py to read your own data format. Check function read_data_input_only in utils.py for an example.