Minimal Neural Atlas

This repository contains the official source code for the ECCV 2022 paper "Minimal Neural Atlas: Parameterizing Complex Surfaces with Minimal Charts and Distortion", which is built on the PyTorch Lightning framework.

Minimal Neural Atlas is a novel explicit neural surface representation that can effectively learn a minimal atlas of 3 charts with distortion-minimal parameterization for surfaces of arbitrary topology. This is enabled by a fully learnable parametric domain, given by an implicit probabilistic occupancy field defined on the $(-1, 1)^2$ open square of the UV parametric space. In contrast, prior works generally predefine the parametric domain, which unnecessarily constrains the boundary and topology of each chart.

If you find our work or code useful, please consider citing us with:

@inproceedings{low2022_mna,
    title={Minimal Neural Atlas: Parameterizing Complex Surfaces with Minimal Charts and Distortion},
    author={Low, Weng Fei and Lee, Gim Hee},
    booktitle={European Conference on Computer Vision (ECCV)},
    year={2022}
}

Installation

We recommend using Conda to set up an environment with the appropriate dependencies for running our code. This can simply be done by running:

git clone https://github.com/wengflow/minimal-neural-atlas.git
cd minimal-neural-atlas
conda env create -f environment.yml

If you prefer a manual installation, the list of dependencies can be found in environment.yml.

Dataset Preparation

Our experiments are conducted on two datasets: CLOTH3D++ and ShapeNet.

CLOTH3D++

1. Download the raw CLOTH3D++ training, validation and test datasets (all tracks)
2. Uncompress the downloaded .tar.gz files and place them in a common folder
3. Preprocess the raw dataset by running the following command from the scripts/ directory:

   python preprocess_cloth3d++.py <raw_dataset_path> <preprocessed_dataset_path>

ShapeNet

1. Download the Occupancy Networks-preprocessed ShapeNet dataset from this link
2. Uncompress the downloaded .zip file to a desired location

Usage

To train, validate or test our model, run the following command from the scripts/ directory:

python run.py {train, val, test} <config_file_path>

Configuration

In the configs/ directory, we provide two pairs of configuration files, {train, test}/surfrec.yaml and {train, test}/svr.yaml, used to train/test our model in the surface and single-view reconstruction experiments, respectively. The specific experimental settings described in the config files are given as follows:

Config File	Experiment	Dataset	No. of Charts	$\mathcal{L}_{dist}$
surfrec.yaml	Surface Reconstruction	CLOTH3D++	2	✓
svr.yaml	Single-View Reconstruction	ShapeNet	3	✗

You should modify the data.dataset_directory, checkpoint_filepath and logger.save_dir parameters in the given configuration files to indicate the path to the preprocessed dataset, pretrained model and your preferred path to the logs, respectively.

To reproduce our reconstruction experiment results under any specific setting, as reported in the paper, you should modify the relevant configuration file provided according to the following guide:

Experimental Setting	Parameter(s) To Be Modified
Dataset	data.dataset, data.dataset_directory
No. of Charts	model.num_charts, {train, eval}_uv_max_sample_size (must be a multiple)
w/ or w/o $\mathcal{L}_{dist}$	loss.weight.distortion

You should also modify checkpoint_filepath and logger.name correspondingly.

Pretrained Models

We provide pretrained Minimal Neural Atlas for both surface and single-view reconstruction experiments at this link. This should allow for a close reproduction of our reported results.