Attentive Implicit Representation Networks (AIR-Nets)
Preprint | Supplementary | Accepted at the International Conference on 3D Vision (3DV)
teaser.mov
This repository is the offical implementation of the paper
AIR-Nets: An Attention-Based Framework for Locally Conditioned Implicit Representations
by Simon Giebenhain and Bastian Goldluecke
Furthermore it provides a unified framework to execute Occupancy Networks (ONets), Convolutional Occuapncy Networks (ConvONets) and IF-Nets.
More qualitative results of our method can be found here.
Install
All experiments with AIR-Nets were run using CUDA version 11.2 and the official pytorch docker image nvcr.io/nvidia/pytorch:20.11-py3
, as published by nvidia here. However, as the model is solely based on simple, common mechanisms, older CUDA and pytorch versions should also work. We provide the air-net_env.yaml
file that holds all python requirements for this project. To conveniently install them automatically with anaconda you can use:
conda env create -f air-net_env.yml
conda activate air-net
AIR-Nets use farthest point sampling (FPS) to downsample the input. Run
pip install pointnet2_ops_lib/.
inorder to install the cuda implementation of FPS. Credits for this go to Erik Wijams's GitHub, from where the code was copied for convenience.
Running
python setup.py build_ext --inplace
installs the MISE algorithm (see http://www.cvlibs.net/publications/Mescheder2019CVPR.pdf) for extracting the reconstructed shapes as meshes.
When you want to run Convolutional Occupancy Networks you will have to install torch scatter
using the official instructions found here.
Data Preparation
In our paper we mainly did experiments with the ShapeNet dataset, but preprocessed in two different falvours. The following describes the preprocessing for both alternatives. Note that they work individually, hence there is no need to prepare both. (When wanting to train with noise I would recommend the Onet data, since the supervision of the IF-Net data is concentrated so close to the boundary that the problem gets a bit ill-posed (adapting noise level and supervision distance can solve this, however).)
Preparing the data used in ONets and ConvONets
To parapre the ONet data clone their repository. Navigate to their repo cd occupancy_networks
and run
bash scripts/download_data.sh
which will download and unpack the data automatically (consuming 73.4 GB). From the perspective of the main repository this will place the data in occupancy_networks/data/ShapeNet
.
Prepating the IF-Net data
A small disclaimer: Preparing the data as in this tutorial will produce ~700GB of data. Deleting the
.obj
and.off
files should reduce the load to 250GB. Storage demand can further be reduced by reducing the number of samples indata_processing/boundary_sampling.py
. If storage is scarce the ONet data (see below) is an alternative.
This data preparation pipeline is mainly copied from IF-Nets, but slightly simplified.
Install a small library needed for the preprocessing using
cd data_processing/libmesh/
python setup.py build_ext --inplace
cd ../..
Furthermore you might need to install meshlab
and xvfb
using
apt-get update
apt-get install meshlab
apt-get install xvfb
To install gcc you can run sudo apt install build-essential
.
To get started, download the preprocessed data by [Xu et. al. NeurIPS'19] from Google Drive into the shapenet
folder.
Please note that some objects in this dataset were made watertight "incorrectly". More specifically some object parts are "double coated", such that the object boundary actually is composed of two boundaries which lie very close together. Therefor the "inside" of such objects lies in between these two boundaries, whereas the "true inside" would be classified as outside. This clearly can lead to ugly reconstructionsl, since representing such a thin "inside" is much trickier.
Then extract the files into shapenet\data
using:
ls shapenet/*.tar.gz |xargs -n1 -i tar -xf {} -C shapenet/data/
Next, the input and supervision data is prepared. First, the data is converted to the .off-format and scaled (such that the longest edge of the bounding box for each object has unit length) using
python data_processing/convert_to_scaled_off.py
Then the point cloud input data can be created using
python data_processing/sample_surface.py
which samples 30.000 point uniformly distributed on the surface of the ground truth mesh. During training and testing the input point clouds will be randomly subsampled from these surface samples. The coordinates and corresponding ground truth occupancy values used for supervision during training can be generated using
python data_processing/boundary_sampling.py -sigma 0.1
python data_processing/boundary_sampling.py -sigma 0.01
where -sigma
specifies the standard deviation of the normally distributed displacements added onto surface samples. Each call will generate 100.000 samples near the object's surface for which ground truth occupancy values are generated using the implicit waterproofing algorithm from IF-Nets supplementary. I have not experimented with any other values for sigma, and just copied the proposed values.
In order to remove meshes that could not be preprocessed correctly (should not be more than around 15 meshes) you should run
python data_processing/filter_corrupted.py -file 'surface_30000_samples.npy' -delete
Pay attantion with this command, i.e. the directory of all objects that don't contain the surface_30000_samples.npy
file are deleted. If you chose to use a different number points, please make sure to adapt the command accordingly.
Finally the data should be located in shapenet/data
.
Preparing the FAUST dataset
In order to download the FAUST dataset visit http://faust.is.tue.mpg.de and sign-up there. Once your account is approved you can download a .zip
-file nameed MPI-FAUST.zip
. Please place the extracted folder in the main folder, such that the data can be found in MPI-FAUST
.
Training
For the training and model specification I use .yaml
files. Their structure is explained in a separate markdown file here, which also has explanations which parameters can tune the model to become less memory intensive.
To train the model run
python train.py -exp_name YOUR_EXP_NAME -cfg_file configs/YOUR_CFG_FILE -data_type YOUR_DATA_TYPE
which stores results in experiments/YOUR_EXP_NAME
. -cfg_file
specifies the path to the config file. The content of the config file will then also be stored in experiments/config.yaml
. YOUR_DATA_TYPE
can either be 'ifnet'
, 'onet'
or 'human'
and dictates which dataset to use. Make sure to adapt the batch_size
parameter in the config file accoridng to your GPU size.
Training progress is saved using tensorboard. You can visualize it by running
tensorboard --logdir experiments/YOUR_EXP_NAME/summary/
Note that checkpoints (including the optimizer) are saved after each epoch in the checkpoints
folder. Therefore training can seamlessly be continued.
Generation
To generate reconstructions of the test set, run
python generate.py -exp_name YOUR_EXP_NAME -checkpoint CKPT_NUM -batch_points 400000 -method REC_METHOD
where CKPT_NUM
specifies the epoch to load the model from and -batch_points
specifies how many points are batched together and may have top be adapted to your GPU size.
REC_METHOD
can either be mise
or mcubes
. The former (and recommended) option uses the MISE algorithm for reconstruciton. The latter uses the vanilla marching cubes algorithm. For the MISE you can specifiy to additional paramters -mise_res
(initial resolution, default is 64) and -mise_steps
(number of refinement steps, defualt 2). (Note that we used 3 refinement steps for the main results of the dense models in the paper, just to be on the save side and not miss any details.) For the regular marching cubes algorithm you can use -mcubes_res
to specify the resolution of the grid (default 128). Note that the cubic scaling quickly renders this really slow.
The command will place the generate meshes in the .OFF
format in experiments/YOUR_EXP_NAME/[email protected]_resxmise_steps/generation
or experiments/YOUR_EXP_NAME/[email protected]_res/generation
depending on method
.
Evaluation
Running
python data_processing/evaluate.py -reconst -generation_path experiments/YOUR_EXP_NAME/evaluation_CKPT.../generation
will evaluate the generated meshes using the most common metrics: the volumetric IOU, the Chamfer distance (L1 and L2), the Normal consistency and F-score.
The results are summarized in experiment/YOUR_EXP_NAME/evaluation_CKPT.../evaluation_results.pkl
by running
python data_processing/evaluate_gather.py -generation_path experiments/YOUR_EXP_NAME/evaluation_CKPT.../generation
Pretrained Models
Weights of trained models can be found here. For example create a folder experiments/PRETRAINED_MODEL
, placing the corresponding config file in experiments/PRETRAINED_MODEL/configs.yaml
and the weights in experiments/PRETRAINED_MODEL/checkpoints/ckpt.tar
. Then run
python generate.py -exp_name PRETRAINED_MODEL -ckpt_name ckpt.tar -data_type DATA_TYPE
Contact
For questions, comments and to discuss ideas please contact Simon Giebenhain via simon.giebenhain (at] uni-konstanz {dot| de.
Citation
@inproceedings{giebenhain2021airnets,
title={AIR-Nets: An Attention-Based Framework for Locally Conditioned Implicit Representations},
author={Giebenhain, Simon and Goldluecke, Bastian},
booktitle={2021 International Conference on 3D Vision (3DV)},
year={2021},
organization={IEEE}
}
Acknowledgements
Large parts of this repository as well as the structure are copied from Julian Chibane's GitHub repository of the IF-Net paper. Please consider also citing their work, when using this repository!
This project also uses libraries form Occupancy Networks by Mescheder et al. CVPR'19 and from Convolutional Occupancy Networks by [Peng et al. ECCV'20].
We also want to thank DISN by [Xu et. al. NeurIPS'19], who provided their preprocessed ShapeNet data publicly. Please consider to cite them if you use our code.
License
Copyright (c) 2020 Julian Chibane, Max-Planck-Gesellschaft and
2021 Simon Giebenhain, Universität Konstanz
Please read carefully the following terms and conditions and any accompanying documentation before you download and/or use this software and associated documentation files (the "Software").
The authors hereby grant you a non-exclusive, non-transferable, free of charge right to copy, modify, merge, publish, distribute, and sublicense the Software for the sole purpose of performing non-commercial scientific research, non-commercial education, or non-commercial artistic projects.
Any other use, in particular any use for commercial purposes, is prohibited. This includes, without limitation, incorporation in a commercial product, use in a commercial service, or production of other artefacts for commercial purposes. For commercial inquiries, please see above contact information.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
You understand and agree that the authors are under no obligation to provide either maintenance services, update services, notices of latent defects, or corrections of defects with regard to the Software. The authors nevertheless reserve the right to update, modify, or discontinue the Software at any time.
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. You agree to cite the Implicit Functions in Feature Space for 3D Shape Reconstruction and Completion
paper and the AIR-Nets: An Attention-Based Framework for Locally Conditioned Implicit Representations
paper in documents and papers that report on research using this Software.