Awesome

Hierarchical Prior Networks

Official PyTorch implementation of the CVPR 2021 paper "Fostering Generalization in Single-view 3D Reconstruction by Learning a Hierarchy of Local and Global Shape Priors". Please cite our paper and the original Occupancy Networks (Mescheder et al.) if you use this work.

[CVPR'21] [Video] [Arxiv]

Overview

Input Depth	Ground Truth
OccNet	HPN (ours)

Installation

You can recreate our anaconda environment mesho3d using

conda env create -f mesho3d.yaml
conda activate mesho3d

Next, compile the extension modules as in the original Occupancy Networks. You can do this via

python setup.py build_ext --inplace

Demo Run

Download the pretrained models from here. Create a folder mkdir out and extract the checkpoints there cd out && unzip pretrained_checkpoints.zip.

Generate meshes for the depth images in data/demo/ using the statements below. You will find them in out/00_paper_genre/p32s16/generation_hierarchical3/ and out/00_paper_genre/p32s16/generation_hierarchical3_sdf/ Note that these networks are trained on airplanes, cars and chairs.

python generate_merge3.py configs/00_paper_genre/genre_bl.yaml configs/00_paper_genre/genre_p32s16.yaml configs/00_paper_genre/genre_p64s32.yaml

python generate_merge3_sdfbl.py configs/00_paper_genre/onet_sdf/genre_bl_sdf.yaml configs/00_paper_genre/genre_p32s16.yaml configs/00_paper_genre/genre_p64s32.yaml

You can visualize the generated meshes .off, .ply and the depth.npz files using:

python vis.py filename

Dataset

Rendering Ground Truth Depth Images

This requires ShapeNetCore to be located in data/

conda env create -f pytorch3d.yaml
source activate pt3d
python render_depth.py

See more options running python render_depth.py --help

Training Data Generation

Rough outline:

• render depth from shapenet shapes in different poses
• we provide the poses, train- and test splits here
• apply the same poses to watertight meshes (see David Stutz)
• sample points from the surface for evaluation
• sample points from the volume (unit cube) and extract occupied / free labels

Training

python train.py configs/00_paper_genre/genre_bl.yaml

Train networks independently the same way as an original Occupancy Network. Use the configs to specify the patch size and stride for the local networks. You can find examples in configs/00_paper_genre/ and configs/00_paper_single/

Generation

python generate_merge3.py configs/00_paper_genre/genre_bl.yaml configs/00_paper_genre/genre_p32s16.yaml configs/00_paper_genre/genre_p64s32.yaml

python generate_merge5.py \
configs/00_paper_single/lamp_bl.yaml \
configs/00_paper_single/lamp_p32s16.yaml \
configs/00_paper_single/lamp_p64s32.yaml \
configs/00_paper_single/lamp_p128s64.yaml \
configs/00_paper_single/lamp_p16s8.yaml

Several versions exist for generate, depending on how many priors should be fused.

Evaluation

This requires the pointcloud.npz file to reside next to the input depth. This file contains 100000 points which have been sampled from the surface of the ground truth mesh in camera space.

python eval_meshes.py configs/config.yaml
python read_results.py configs/config.yaml

Citation

If you find our paper or this code useful, please cite:

@inproceedings{Bechtold2021HPN,
    title = {Fostering Generalization in Single-view 3D Reconstruction by Learning a Hierarchy of Local and Global Shape Priors},
    author = {Bechtold, Jan and Maxim, Tatarchenko and Volker, Fischer and Thomas, Brox},
    booktitle = {CVPR},
    year = {2021}
}

We build upon the code of Occupancy Networks (Mescheder et al. CVPR 2019). If you use our work please also cite:

@inproceedings{Occupancy Networks,
    title = {Occupancy Networks: Learning 3D Reconstruction in Function Space},
    author = {Mescheder, Lars and Oechsle, Michael and Niemeyer, Michael and Nowozin, Sebastian and Geiger, Andreas},
    booktitle = {Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR)},
    year = {2019}
}

Many thanks to them for this great work and great code.

License

Hierarchical Prior Networks is open-sourced under the AGPL-3.0 license. See the LICENSE.md file for details. For a list of other open source components included in HierarchicalPriorNetworks, see the file 3rd-party-licenses.txt.