🚀 Major upgrade 🚀 : Migration to Pytorch v1 and Python 3.7. The code is now much more generic and easy to install.

AtlasNet

This repository contains the source codes for the paper AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation . The network is able to synthesize a mesh (point cloud + connectivity) from a low-resolution point cloud, or from an image.

Citing this work

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

@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}
        }

Project Page

The project page is available http://imagine.enpc.fr/~groueixt/atlasnet/

Install

Clone the repo and install dependencies

This implementation uses Pytorch.

## Download the repository
git clone https://github.com/ThibaultGROUEIX/AtlasNet.git
cd AtlasNet
## Create python env with relevant packages
conda create --name pytorch-atlasnet python=3.7
source activate pytorch-atlasnet
pip install pandas visdom
conda install pytorch torchvision -c pytorch # or from sources if you prefer
# you're done ! Congrats :)

Tested on 11/18 with pytorch 0.4.1 (py37_py36_py35_py27__9.0.176_7.1.2_2) and latest source

Demo

Require 3GB RAM on the GPU and 5sec to run. Pass --cuda 0 to run without gpu (9sec).

cd trained_models; ./download_models.sh; cd .. # download the trained models
python inference/demo.py --cuda 1

This script takes as input a 137 * 137 image (from ShapeNet), run it through a trained resnet encoder, then decode it through a trained atlasnet with 25 learned parameterizations, and save the output to output.ply

Training

Data

cd data; ./download_data.sh; cd ..

We used the ShapeNet dataset for 3D models, and rendered views from 3D-R2N2:

When using the provided data make sure to respect the shapenet license.

• The point clouds from ShapeNet, with normals go in data/customShapeNet
• The corresponding normalized mesh (for the metro distance) go in data/ShapeNetCorev2Normalized
• the rendered views go in data/ShapeNetRendering

The trained models and some corresponding results are also available online :

• The trained_models go in trained_models/

In case you need the results of ICP on PointSetGen output :

• ICP on PSG

Build chamfer distance (optional)

Using the custom chamfer distance will divide memory usage by 2 and will be a bit faster. Use it if you're short on memory especially when training models for Single View reconstruction.

source activate pytorch-atlasnet
cd ./extension
python setup.py install

Start training

• First launch a visdom server :

python -m visdom.server -p 8888

• Launch the training. Check out all the options in ./training/train_AE_AtlasNet.py .

export CUDA_VISIBLE_DEVICES=0 #whichever you want
source activate pytorch-atlasnet
git pull
env=AE_AtlasNet
nb_primitives=25
python ./training/train_AE_AtlasNet.py --env $env --nb_primitives $nb_primitives |& tee ${env}.txt

• Monitor your training on http://localhost:8888/

• Compute some results with your trained model

  python ./inference/run_AE_AtlasNet.py

  The trained models accessible here have the following performances, slightly better than the one reported in the paper. The number reported is the chamfer distance.

Method	Chamfer⁽⁰⁾	GPU memory⁽¹⁾	Time by epoch⁽²⁾
Autoencoder_Atlasnet_25prim	0.0014476474650672833	4.1GB	6min55s
Autoencoder_Atlasnet_1sphere	0.0017207141972321953	3.6GB	5min30s
Autoencoder_Baseline	0.001963350556556298	1.9GB	2min05s
SingleViewReconstruction_Atlasnet_25prim	0.004638490150569042	6.8GB	10min04s
SingleViewReconstruction_Atlasnet_1sphere	0.005198702077052366	5.6GB	8min16s
SingleViewReconstruction_Baseline	0.0048062904884797605	1.7GB	3min30s

⁽⁰⁾ computed between 2500 ground truth points and 2500 reconstructed points.

⁽¹⁾ with the flag --accelerated_chamfer 1.

⁽²⁾this is only an estimate, the code is not optimised. The easiest way to enhance it would be to preload the training data to use the GPU at 100%. Time computed with the flag --accelerated_chamfer 1.

Autoencoder : 25 learned parameterization

val_loss	watercraft	monitor	car	couch	cabinet	lamp	plane	speaker	table	chair	bench	firearm	cellphone
0.0014795344685297⁽³⁾	0.00127737027906	0.0016588120616	0.00152693425022	0.00171516126198	0.00168296881168	0.00232362473947	0.000833268054194	0.0025417242402	0.00149979386376	0.00156113364435	0.00120812499892	0.000626943988977	0.0012117530635

Single View Reconstruction : 25 learned parameterization

val_loss	watercraft	monitor	car	couch	cabinet	lamp	plane	speaker	table	chair	bench	firearm	cellphone
0.00400863720389⁽³⁾	0.00336707355723	0.00456469316226	0.00306795421868	0.00404269965806	0.00355917039209	0.0114094304694	0.00192791500002	0.00780984506137	0.00368373458016	0.00407004468516	0.0030023689528	0.00192803189235	0.00293665724291

⁽³⁾the number is slightly different for above because it comes from legacy code (Atlasnet v1).

• Evaluate quantitatively the reconstructed meshes : METRO DISTANCE

Visualisation

The generated 3D models' surfaces are not oriented. As a consequence, some area will appear dark if you directly visualize the results in Meshlab. You have to incorporate your own fragment shader in Meshlab, that flip the normals in they are hit by a ray from the wrong side. An exemple is given for the Phong BRDF.

sudo mv /usr/share/meshlab/shaders/phong.frag /usr/share/meshlab/shaders/phong.frag.bak
sudo cp auxiliary/phong.frag /usr/share/meshlab/shaders/phong.frag #restart Meshlab

Acknowledgement

The code for the Chamfer Loss was adapted from Fei Xia' repo : PointGan. Many thanks to him !

This work was funded by Adobe System and Ecole Doctorale MSTIC.

Cool Contributions

• Yana Hasson trained our sphere model, for Single View Reconstruction (SVR) in view-centered coordinates : performances are unaffected! Qualitative and quantitative results follow. Many thanks ! View this paper for a good review of on object-centered representation VS view-centered representation.

frame	Average recontruction error for SVR (x1000) : chamfer distance on input pointcloud and reconstruction of size 2500 pts
object-centered	4.87⁽⁴⁾
view-centered	4.88

⁽⁴⁾ Trained with Atlasnet v2 (with learning rate scheduler : slightly better than the paper's result)

License

MIT

Old links for data

• The point clouds from ShapeNet, with normals go in data/customShapeNet
• The corresponding normalized mesh (for the metro distance) go in data/ShapeNetCorev2Normalized
• the rendered views go in data/ShapeNetRendering

The trained models and some corresponding results are also available online :

• The trained_models go in trained_models/