Awesome

Neural Semantic Surface Maps (NSSM)

This repository contains code to train a Neural Semantic Surface Maps and a Neural Surface Maps.

Setup
Preprocessing
Optimization
Bibtex

Setup <a name="setup"/>

To set up the environment you can download the docker image and build it or install all the required packages in your conda environment (see docker image).

docker build --build-arg username=luca --build-arg userid=`id -u` -t nssm ./nssm/

This will create a docker image, the sudo password is docker.

Once you have the docker image up and running, run the docker container and link it to your workspace to access the code.

docker run --security-opt seccomp=unconfined -h DOCKER --name nssm  -v ~/workspace:/home/luca/workspace --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=0 -it --shm-size=2gb nssm:latest bash

Please adjust the path to your workspace!

NOTE: the instructions above create a user called luca and mount the workspace in the user folder, if you specify a different username then change it also when running the docker container.

This configuration was tested on Ubuntu 20.04 with a 2080Ti.

Now compile the C++ code to parametrize a mesh:

cd parametrization_cpp
mkdir build
cd build
cmake ..
make -j4

Preprocessing <a name="preprocess"/>

Create a folder for the shape pair, inside it create a folder called meshes containing the source and target mesh (both obj files).

cd workspace
mkdir nssm_pair
mkdir meshes
cp source.obj nssm_output/meshes/source.obj
cp target.obj nssm_output/meshes/target.obj

That's it, you are ready to optimize a map between them!

NOTE: the code will look for folder called meshes containing a source.obj and target.obj file. If this does not happen, then the whole pipeline will crash.

Optimization <a name="optimize"/>

You can run the full pipeline with just one script:

cd nssm
./run_pair.sh ~/workspace/nssm_pair

Remember, the path must contain a meshes folder as described in the preprocessing step. The code will automatically generate and save inside the folder nssm_pair the output data for each stage. The map is contained inside the nssm_pair/map folder (both model weights and meshes).

Bibtex <a name="bibtex"/>

@article{morreale2024neural,
    title={Neural Semantic Surface Maps},
    author={Morreale, Luca and Aigerman, Noam and Kim, Vladimir G. and Mitra, Niloy J.},
    booktitle={Computer Graphics Forum},
    volume={43},
    number={2},
    year={2024},
    organization={Wiley Online Library}
}