AtlasNet + Shape Reconstruction by Learning Differentiable Surface Representations

Welcome to this branch : it is an unofficial implemention of Shape Reconstruction by Learning Differentiable Surface Representations from Jan Bednarik and Thibault Groueix.

Note

This repo include an implementation from Jan Bednarik of Differentiable Surface properties and their usage to form a conformal regularization of 3D shape reconstruction (on ShapeNet).

Such properties include (please check Jan's paper for details):

• Normals
• First fondamental form (Jacobian)
• Mean curvature
• Gaussian curvature

Install

This implementation uses Python 3.6, Pytorch, Pymesh, Cuda 10.1.

# Copy/Paste the snippet in a terminal
git clone --recurse-submodules https://github.com/ThibaultGROUEIX/AtlasNet.git
cd AtlasNet 

#Dependencies
conda create -n atlasnet python=3.6 --yes
conda activate atlasnet
conda install  pytorch torchvision cudatoolkit=10.1 -c pytorch --yes
pip install --user --requirement  requirements.txt # pip dependencies

Optional : Compile Chamfer (MIT) + Metro Distance (GPL3 Licence)

# Copy/Paste the snippet in a terminal
python auxiliary/ChamferDistancePytorch/chamfer3D/setup.py install #MIT
cd auxiliary
git clone https://github.com/ThibaultGROUEIX/metro_sources.git
cd metro_sources; python setup.py --build # build metro distance #GPL3
cd ../..

Training

Same options as the main branch. See here Training Just add --conformal_regul --lambda_conformal_regul 0.001 to train with Differentiable Surface properties.

Autoencoder [default]

python train.py --shapenet13 --dir_name log/atlasnet_autoencoder_25_squares --nb_primitives 25 --template_type "SQUARE" --conformal_regul --lambda_conformal_regul 0.001

Single-View Reconstruction [default]

python train.py --shapenet13 --dir_name log/atlasnet_singleview_25_squares_tmp --nb_primitives 25 --template_type SQUARE --SVR --reload_decoder_path log/atlasnet_autoencoder_25_squares --train_only_encoder --conformal_regul --lambda_conformal_regul 0.001

🖐️ Monitor your training on http://localhost:8890/

🖐️ See report on a completed training at http://localhost:8891/{DIR_NAME}

Using a trained model

python train.py --demo --demo_input_path YOUR_IMAGE_or_OBJ_PATH --reload_model_path YOUR_MODEL_PTH_PATH

This function takes an image or pointcloud path as input and save the mesh infered by Atlasnet
Extension supported are `ply` `npy` `obg` and `png`
--demo_input_path input file e.g. image.png or object.ply 
--reload_model_path trained model path (see below for pretrained models) 
:return: path to the generated mesh

Quantitative Results Autoencoder Pretrained models

Method (on Planes)	Chamfer (*1)	Fscore (*2)	Metro (*3)	Total Train time (min)
Autoencoder 1 Square	1.40	85.3	19	30
Autoencoder 1 Square + Conformal (lamda=1)	2.8	72.9	24	300
Autoencoder 1 Square + Conformal (lamda=0.001)	1.45	85.3	20	300

• (*1) x1000. Computed between 2500 ground truth points and 2500 reconstructed points.
• (*2) The threshold is sqrt(0.001)
• (*3) x100. Metro is ran on unormalized point clouds (which explains a difference with the paper's numbers)

Citing this work

@article{bednarik2019shape,
   title={Shape Reconstruction by Learning Differentiable Surface Representations},
 	 author={Bednarik, Jan and Parashar, Shaifali and Gundogdu, Erhan and Salzmann, Mathieu and Fua, Pascal},
   booktitle={Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR)},
 	 year={2019}
}

@inproceedings{groueix2018,
          title={{AtlasNet: A Papier-M\^ach\'e Approach to Learning 3D Surface Generation}},
          author={Groueix, Thibault and Fisher, Matthew and Kim, Vladimir G. and Russell, Bryan and Aubry, Mathieu},
          booktitle={Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR)},
          year={2018}
        }