I adapted this code to use pytorch 2.2.1 and pytorch-lighting 2.2.1. It runs well on RTX4090. I basically follow the instructions in the original repo, except that build torch-gpu-svd with CUDA-12.0.

Official implementation of Neural Surface Maps - Luca Morreale, Noam Aigerman, Vladimir Kim, Niloy J. Mitra

[Paper] [Project Page]

Replicating the results is possible following these steps:

1. Parametrize the surface
2. Prepare surface sample
3. Overfit the surface
4. Neural parametrization of the surface
5. Optimize surface-to-surface map
6. Optimize a map between a collection

This is a preprocessing step. You can use SLIM[1] from this repo to fulfill this step.

Given a parametrized surface (prev. step), we need to convert it into a sample. First of all, we need to over sample the surface with Meshlab. You can use the midpoint subdivision filter.

Once the super-sampled surface is ready then you can convert it into a sample:

python -m preprocessing.convert_sample surface_slim.obj surface_slim_oversampled.obj output_sample.pth

The file output_sample.pth is the sample ready to be over-fitted.

A surface representation is generated with:

python -m training_surface_map dataset.sample_path=output_sample.pth

This will save a surface map inside outputs/neural_maps folder. The folder name follows this patterns: overfit_[timestamp]. Inside that folder, the map is saved under the sample fodler as pth file.

The overfitted surface can be generated with:

python -m show_surface_map

please, set the path to the pth file just created inside the script.

Generating a neural parametrization need to run:

python -m training_parametrization_map dataset.sample_path=your_surface_map.pth

Like for the overfitting, this saves the map inside outputs/neural_maps folder. The folder name have the following patterns parametrization_[timestamp].

To display the paramtrization obtained run:

python -m show_parametrization_map

please, set the path to the pth file just created inside the script.

To generating a inter-surface map run:

python -m training_intersurface_map dataset.sample_path_g=your_surface_map_a.pth dataset.sample_path_f=your_surface_map_b.pth

Note, this steps requires two surface maps. A source, sample_path_g, and a target, sample_path_f.

Likewise the overfitting, the map is saved inside outputs/neural_maps. The inter-surface map folder pattern is intersurface_[timestamp]. The pth file is inside the models folder.

To display the inter-surface map run:

python -m show_intersurface_map

remember to set the path of the maps inside the script.

A collection between a set of surface maps can be optimized with:

python -m training_intersurface_map dataset.sample_path_g=your_surface_map_g.pth dataset.sample_path_f=your_surface_map_f.pth dataset.sample_path_q=your_surface_map_q.pth

Note, this steps requires three surface maps. A source, sample_path_g, and two targets, sample_path_f and sample_path_q.

This will save two maps inside outputs/neural_maps folder. The folder name follows this patterns: collection_[timestamp], under the folder models you can find two *.pth file.

To display the collection map run:

python -m show_collection_map

remember to set the path of maps inside the script.

Dependencies are listed in environment.yml. Using conda, all the packages can be installed with conda env create -f environment.yml.

On top of the packages above, please install also pytorch svd on gpu package.

Any mesh can be used for this process. A data example can be downloaded here.

@misc{morreale2021neural,
      title={Neural Surface Maps},
      author={Luca Morreale and Noam Aigerman and Vladimir Kim and Niloy J. Mitra},
      year={2021},
      eprint={2103.16942},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

[1] Scalable locally injective mappings - Michael Rabinovich et. al. - ACM Transactions on Graphics (TOG) 2017