ThibaultGROUEIX/AtlasNet

AtlasNet [Project Page] [Paper] [Talk]

AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation
Thibault Groueix, Matthew Fisher, Vladimir G. Kim , Bryan C. Russell, Mathieu Aubry
In CVPR, 2018.

:rocket: New branch : AtlasNet + Shape Reconstruction by Learning Differentiable Surface Representations

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 ../..

Usage

Demo : python train.py --demo
Training : python train.py --shapenet13 Monitor on http://localhost:8890/
Latest Refacto 12-2019
- [x] Factorize Single View Reconstruction and autoencoder in same class
- [x] Factorise Square and Sphere template in same class
- [x] Add latent vector as bias after first layer(30% speedup)
- [x] Remove last th in decoder
- [x] Make large .pth tensor with all pointclouds in cache(drop the nasty Chunk_reader)
- [x] Make-it multi-gpu
- [x] Add netvision visualization of the results
- [x] Rewrite main script object-oriented
- [x] Check that everything works in latest pytorch version
- [x] Add more layer by default and flag for the number of layers and hidden neurons
- [x] Add a flag to generate a mesh directly
- [x] Add a python setup install
- [x] Make sure GPU are used at 100%
- [x] Add f-score in Chamfer + report f-score
- [x] Get rid of shapenet_v2 data and use v1!
- [x] Fix path issues no more sys.path.append
- [x] Preprocess shapenet 55 and add it in dataloader
- [x] Make minimal dependencies

Quantitative Results

Method	Chamfer (*1)	Fscore (*2)	Metro (*3)	Total Train time (min)
Autoencoder 25 Squares	1.35	82.3%	6.82	731
Autoencoder 1 Sphere	1.35	83.3%	6.94	548
SingleView 25 Squares	3.78	63.1%	8.94	1422
SingleView 1 Sphere	3.76	64.4%	9.01	1297

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

Related projects

Learning Elementary Structures
3D-CODED
Cycle Consistent Deformations
Atlasnet code V2.2 (more linear, script like, may be easier to understand at first)

Citing this work

@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}
        }