DefTet

This is the official repo for:

Learning Deformable Tetrahedral Meshes for 3D Reconstruction (NeurIPS 2020)

Jun Gao, Wenzheng Chen, Tommy Xiang, Clement Fuji Tsang, Alec Jacobson, Morgan McGuire, Sanja Fidler

NeurIPS 2020, Paper, Supplementary, Project Page,

Requirements

• Python 3.8 is supported.
• Pytorch 1.9.0.
• This code is tested with CUDA 11.1.
• GCC >= 6.0

Install Pytorch

pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 -f https://download.pytorch.org/whl/torch_stable.html

Install QuarTet following official Link.

git clone https://github.com/crawforddoran/quartet
apt-get install libglfw3-dev libgl1-mesa-dev libglu1-mesa-dev freeglut3-dev # Dependencies for QuarTet
cd quartet
make depend
make 

Install Kaolin following official Link.

Install other required libraries by doing:

apt-get update
apt-get install tmux htop -y
pip install opencv-python tensorboardx meshzoo ipdb imageio
apt-get install ffmpeg libsm6 libxext6  -y
HOME_DIR="$PWD"
cd "$HOME_DIR/utils/lib/tet_adj_share"
bash do_all.sh
cd "$HOME_DIR/utils/lib/tet_face_adj"
bash do_all.sh
cd "$HOME_DIR/utils/lib/tet_point_adj"
bash do_all.sh
cd "$HOME_DIR/utils/lib/colaps_v"
bash do_all.sh
cd $HOME_DIR

Training

Dataset

We use ShapeNet Core dataset, please first download ShapeNet following Official Link into the directory /data/shapenet_kaolin, then inside the code, we will use kaolin to preprocess the dataset (see details at dataloader.py)

Training on 3D Point Cloud Reconstruction

We experimented training on different resolution by passing the config --res xxx. To reproduce the results in paper, we recommond using res 70, training at higher resolution won't have too much improvement.

python train_multigpu.py  --pow 4 --save_vis  --batch_size 8 --print_every 300  --dataset_dir /data/shapenet_kaolin  --save_vis_ever 10000  --no_use_pos_encoding --no_use_vert_feat --use_init_pos_mask --point_cloud --lambda_surf 5 --lambda_surf_chamfer 1 --lambda_amips 1 --res 70 --no_expand_boundary --use_two_encoder --no_use_vert_feat --no_use_pvcnn_pos_decoder --no_use_dvr_pos_decoder --use_gcn_pos_decoder --no_use_dvr_occ_decoder --add_input_noise --use_pvcnn_occ_decoder  --use_all --experiment_id pc_pvcnn_gcn_70_surf_1_chamfer_1_all_scale_pvcnn --scale_pvcnn

Inference on 3D Point Cloud Reconstruction using trained model

This evaluation scripts provides the results on different metrics, including chamfer, chamfer L1, F-score, Hausdorff distance.

python eval.py ----experiment_path experiments/YOUR-EXP-NAME

Optimization that uses 2D supervision

Here we provide the optimization script on using our diff render for dmtet to optimize a tet mesh using volume rendering.

cd diff_render/diftet_6_subdiv/6_optim
pip install configargparse
python optim_with_mask_subdiv_from_gridmov.py --expname hotdog --datadir YOUR_NERF_DIR --savedir YOUR_SAVING_DIR --remote

Ciatation

If you use the code, please cite our paper:

@inproceedings{gao2020deftet,
title={Learning Deformable Tetrahedral Meshes for 3D Reconstruction},
author={Jun Gao and Wenzheng Chen and Tommy Xiang and Clement Fuji Tsang and Alec Jacobson and Morgan McGuire and Sanja Fidler},
booktitle={Advances In Neural Information Processing Systems},
year={2020}
}

Following works

We have some following works related to this one:

• Deep Marching Tetrahedra: a Hybrid Representation for High-Resolution 3D Shape Synthesis. NeurIPS 2021
• Extracting Triangular 3D Models, Materials, and Lighting From Images. (CVPR) 2022

Acknowledgement

The code for PVCNN module in this codebase is borrowed from PVCNN (MIT License)