Home

Awesome

NeAT

This is the official repo for the implementation of NeAT: Learning Neural Implicit Surfaces with Arbitrary Topologies from Multi-view Images. alt text

Project page | Paper | Data & Supp

Usage

Setup the environment

You can create an anaconda environment called neat_env by running the following commands:

git clone https://github.com/xmeng525/NeAT.git
cd NeAT
conda create -n neat_env python=3.8
conda activate neat_env
pip install -r requirements.txt

Demo

You can now test our code on the provided input images and checkpoints.

For example, run the following command to reconstruct a T-shirt from the MGN dataset:

python exp_runner.py --case MGN/TShirtNoCoat_125611500935128 --conf ./confs/wmask.conf --is_continue --mode validate_mesh --res 256

This script takes the images in data/MGN/TShirtNoCoat_125611500935128 as the input, and loads the pretrained checkpoint in exp/MGN/TShirtNoCoat_125611500935128/checkpoints.

The generated meshes are saved in exp/MGN/TShirtNoCoat_125611500935128/meshes:

Welcome to download more dataset and checkpoints from HERE.

Similar to the example above, please throw the downloaded data into data/DATASET_NAME and throw the corresponding checkpoint into exp/DATASET_NAME/CASE_NAME/wmask/checkpoints.

Note:

For the data data_DTU/*, please use --conf ./confs/wmask_dtu.conf.

For the data others/cat_mask, please use --conf ./confs/wmask_onlypos.conf.

Running

Method 1: Please refer to the Data Conversion in NeuS to generate your own data.

Method 2: If you plan to generate image & cameras used for NeAT reconstruction from a mesh or point cloud, you may checkout ./data_generation.

With the images and masks ready, you may reconstruct neural implicit surface with arbitrary topologies from multi-view images by running:

python exp_runner.py --case YOUR_CASE --conf ./confs/wmask.conf

In the evaluation stage, you may export the mesh reconstructed with the multi-view images by running:

python exp_runner.py --case YOUR_CASE --conf ./confs/wmask.conf --is_continue --mode validate_mesh --res 512

Citation

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

@InProceedings{Meng_2023_CVPR,
    author    = {Meng, Xiaoxu and Chen, Weikai and Yang, Bo},
    title     = {NeAT: Learning Neural Implicit Surfaces With Arbitrary Topologies From Multi-View Images},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {248-258}
}

Acknowledgement

Some code snippets are borrowed from: IDR, NeRF-pytorch, and NeuS. Thanks for these great projects!