[arXiv] [BibTex]

3DQD: Generalized Deep 3D Shape Prior via Part-Discretized Diffusion Process
Yuhan Li, Yishun Dou, Xuanhong Chen, Bingbing Ni, Yilin Sun, Yutian Liu and Fuzhen Wang

[abstract] 3DQD is a generalized 3D shape generation prior model, tailored for multiple 3D tasks including unconditional shape generation, point cloud completion, and crossmodality shape generation, etc. 3DQD first learns a compact representation with P-VQ-VAE for its advantages in computational saving and consistency among different tasks. Then a novel discrete diffusion generator is trained with accurate, expressive and diversified object structural modeling. Multi-frequency fusion modules are developed to suppress high-frequency outliers.

We have released the demo to show extended applications of single-view reconstruction!

Please refer to demo_imgdf.ipynb.

• Python >= 3.8
• CUDA 11.1
• Pytorch >= 1.9
• Pytorch3D
• trimesh
• tqdm
• scipy
• tensorboard
• PyMCubes

Or you can setup the environment using conda:

conda env create -f environment.yaml
conda activate 3dqd

(Optional) Some packages are also need to evaluate generated shapes.

• clip >= 1.0 (The pretrained clip model ViT-B-32.pt can be download here)
• light-field-distance == 0.0.9
• Frechet Pointcloud Distance (Following the instructions from TreeGAN)
• Pytorch Chamfer Distance

cd metrics/pvd_metrics/ChamferDistancePytorch/chamfer3D
python setup.py install

• PyTorchEMD

cd metrics/pvd_metrics/PyTorchEMD
python setup.py install
cp build/**/emd_cuda.cpython-38-x86_64-linux-gnu.so .

• pointnet2-ops

pip install git+https://gitee.com/Fukexue/Pointnet2_PyTorch.git@acda965224f35854bc331cd5fe140393216b0a71#subdirectory=pointnet2_ops_lib

Pre-trained models are available here.
The download the pre-trained models from the provided links should be put into the ./saved_ckpt folder.

[One Drive] We follow the instructions from AutoSDF to preprocess data.

1. ShapeNet

First you need to download the ShapeNetCore.v1 following the instruction of https://www.shapenet.org/account/.

To extract SDF values, we followed the preprocessing steps from DISN.

The preprocessed data can be downloaded from One Drive

2. ShapeGlot

cd {dataroot}   # Here is your own Dataroot
git clone https://github.com/optas/shapeglot

# Follow the instructions from ShapeGlot to download the shape-text pairs data.
# You need first finish a questionnaire in https://github.com/optas/shapeglot. 
# Then you will receive a link and put it into the file 'download_data.sh'.

cd shapeglot/
./download_data.sh

3. (Optional for single-view reconstruciton)Pix3D

We preprocess the Pix3D datasets similarly to ShapeNet.

The preprocessed data can be downloaded from One Drive.

The layout should look like this

├── 3DQD
├── dataset [This is your dataroot]
│   ├── ShapeGlot
│   ├── pix3d
│   ├── ShapeNet
│   │   ├── 20691156
│   │   │   ├── *****************
│   │   │   │   ├── ori_sample.h5
│   │   ├── 20747177
│   │   ├── filelists
│   │   │   ├── 20691156_test.lst
│   │   │   ├── 20691156_train.lst
│   │   ├── SDF_v1_64
│   │   │   ├── 20691156
│   │   │   │   ├── *****************
│   │   │   │   │   ├── ori_sample.h5       
│   ├── extracted_code [This folder will be generated after excute ./launchers/extract_pvqvae_snet.sh]
│   |   ├── pvqvae-snet-all-T0.2
│   |   |   ├── 20691156
│   |   │   │   ├── *****************
│   |   │   │   │   ├── code.npy
│   |   │   │   │   ├── codeix.npy
│   |   │   │   │   ├── x.npy

We offer three kinds of scripts to evaluate the trained model. Every script will produce a logger log.txt to record your hyperparameter settings and metrics during evaluating.

# To evaluating models for unconditional generation.
python evaluation_scripts/eval_uncond.py --cat {car|airplane|chair} --ckpt {MODEL_PATH}

# To evaluating models for shape completion.
python evaluation_scripts/eval_shape_comp.py --ckpt {MODEL_PATH} --comp_type {half|octant}

# To evaluating models for text-driven generation.
python evaluation_scripts/eval_textdf.py --ckpt {MODEL_PATH}

The whole training has two stages, which means the P-VQ-VAE and discrete diffusion generator are trained seperately.

1. First train the P-VQ-VAE on ShapeNet:

./launchers/train_pvqvae_snet.sh

After training, copy the trained P-VQ-VAE checkpoint (pretrained-vqvae-snet.ckpt) to the ./saved_ckpt folder. Based on this quantized feature representation of shape, diffusion models on various tasks can be trained.

2. Then extract the code for each sample of ShapeNet (caching them for training the diffusion):

./launchers/extract_pvqvae_snet.sh

3. Train the diffusion models for various tasks:

 # Tarin the discrete diffusion generator on 13 categories in Shapenet for shape completion.
./launchers/train_comp_df_snet_code.sh  

# Tarin the discrete diffusion generator on single(car|chair|airplane) category for unconditional generation.
./launchers/train_uncond_df_snet_code.sh

# Tarin the discrete diffusion generator on chair category for text-driven generation.
./launchers/train_textdf_shapeglot.sh

If you find our work useful to your research, please consider citing:

@misc{2303.10406,
Author = {Yuhan Li and Yishun Dou and Xuanhong Chen and Bingbing Ni and Yilin Sun and Yutian Liu and Fuzhen Wang},
Title = {3DQD: Generalized Deep 3D Shape Prior via Part-Discretized Diffusion Process},
Year = {2023},
Eprint = {arXiv:2303.10406},
}

This code borrowed heavily from AutoSDF, VQ-DIFF and LDM. Thanks for their great work!