Skip to content

code for paper "Learning to Generate 3D Shapes from a Single Example", SIGGRAPH Asia 2022

License

Notifications You must be signed in to change notification settings

ChrisWu1997/SingleShapeGen

Repository files navigation

Learning to Generate 3D Shapes from a Single Example

teaser

Official implementation for the paper:

Learning to Generate 3D Shapes from a Single Example
Rundi Wu, Changxi Zheng
Columbia University
SIGGRAPH Asia 2022 (Journal Track)

Installation

Prerequisites:

  • python 3.9+
  • An Nvidia GPU

Install dependencies with conda:

conda env create -f environment.yml
conda activate ssg

or install dependencies with pip:

pip install -r requirement.txt
# NOTE: check https://pytorch.org/ for pytorch installation command for your CUDA version

Pretrained models

We provide pretrained models for all shapes that are used in our paper. Download all of them (about 1G):

bash download_models.sh

or download each model individually, e.g.,

bash download_models.sh ssg_Acropolis_r256s8

Backup Google Drive link.

Quick start: a GUI demo

We provide a simple GUI demo (based on Open3D) that allows quick shape generation with a trained model. For example, run

python gui_demo.py checkpoints/ssg_Acropolis_r256s8

ssg_demo

(Recorded on a Ubuntu 20.04 with an NVIDIA 3090 GPU. Also tested on a Window 11 with an NVIDIA 2070 GPU.)

Inference

Random generation

To randomly generate new shapes, run

python main.py test --tag ssg_Acropolis_r256s8 -g 0 --n_samples 10 --mode rand

The generated shapes will be saved in .h5 format, compatible with the training data.

Specify --resize to change the spatial dimensions. For example, --resize 1.5 1.0 1.0 generates shapes whose size along x-axis are 1.5 times larger than original.

Specify --upsample to construct the output shape at a higher resolution. For example, --upsample 2 gives in 2 times higher resolution.

Interpolation and extrapolation

For interpolation and extrapolation between two randomly generated samples, run

python main.py test --tag ssg_Acropolis_r256s8 -g 0 --n_samples 5 --mode interp

Visualize and export

To quickly visualize the generated shapes (of .h5 format), run

python vis_export.py -s checkpoints/ssg_Acropolis_r256s8/rand_n10_bin_r1x1x1 -f mesh --smooth 3 --cleanup

--smooth specifies Laplacian smoothing iterations. --cleanup keeps only the largest connected component.

Specify --export obj to export meshes in obj format.

Training data preparation

We list the sources for all example shapes that we used: data/README.md. Most of them are free and you can download accordingly.

To construct the training data (voxel pyramid) from a mesh, we rely on binvox. After downloading it, make sure you change BINVOX_PATH in voxelization/voxelize.py to your path to excetuable binvox. Then run our script

cd voxelization
python voxelize.py -s {path-to-your-mesh-file} --res 128 --n_scales 6 -o {save-path.h5}
# --res: finest voxel resolution. --n_scales: number of scales.

The processed data will be saved in .h5 format.

TBA: release preprocessed data?

Training

To train on the processed h5 data, run

python main.py train --tag {your-experiment-tag} -s {path-to-processed-h5-data} -g {gpu-id}

By default, the log and model will be saved in checkpoints/{your-experiment-tag}.

Evaluation

We provide code for evaluation metrics LP-IoU, LP-F-score, SSFID and Div. SSFID relies on a pretrained 3D shape classifier. Please download it from here and put Clsshapenet_128.pth under evaluation folder.

To perform evaluation, we first randomly generate 100 shapes, e.g.,

python main.py test --tag ssg_Acropolis_r128s6 -g 0 --n_samples 100 --mode rand

Then run the evalution script to compute all metrics, e.g.,

cd evaluation
# ./eval.sh {generated-shapes-folder} {reference-shape} {gpu-ids}
./eval.sh ../checkpoints/ssg_Acropolis_r128s6/rand_n100_bin_r1x1x1 ../data/Acropolis_r128s6.h5 0

See evaluation folder for evalution scripts for each individual metric.

SinGAN-3D baseline

We also provide a baseline, SinGAN-3D, as described in our paper. To use it, simply specify --G_struct conv3d when training the model. Pretrained models are also provided (begin with "singan3d").

Rendering

We provide code and configurations for rendering figures in our paper. We rely on Blender and BlenderToolbox. To use our rendering script, make sure have them installed and change the corresponding paths (BLENDER_PATH and BLENDERTOOLBOX_PATH in render/blender_utils.py). Then run

cd render
python render_script.py -s {path-to-generated-shapes-folder} -c {render-config-name} --smooth 3 --cleanup

See render/render_configs.json for saved rendering configs.

Acknowledgement

We develop some part of this repo based on code from SinGAN, ConSinGAN, DECOR-GAN and BlenderToolbox. We would like to thank their authors.

Citation

@article{wu2022learning,
    title={Learning to Generate 3D Shapes from a Single Example},
    author={Wu, Rundi and Zheng, Changxi},
    journal={ACM Transactions on Graphics (TOG)},
    volume={41},
    number={6},
    articleno={224},
    numpages={19},
    year={2022},
    publisher={ACM New York, NY, USA}
}