Ghost on the Shell: An Expressive Representation of General 3D Shapes

Introduction

This is the official implementation of our paper (ICLR 2024 oral) "Ghost on the Shell: An Expressive Representation of General 3D Shapes" (G-Shell).

G-Shell is a generic and differentiable representation for both watertight and non-watertight meshes. It enables 1) efficient and robust rasterization-based multiview reconstruction and 2) template-free generation of non-watertight meshes.

Please refer to our project page and our paper for more details.

Getting Started

Requirements

• Python >= 3.8
• CUDA 11.8
• PyTorch == 1.13.1

(Conda installation recommended)

Reconstruction

Run the following

pip install ninja imageio PyOpenGL glfw xatlas gdown
pip install git+https://github.com/NVlabs/nvdiffrast/
pip install --global-option="--no-networks" git+https://github.com/NVlabs/tiny-cuda-nn#subdirectory=bindings/torch

Follow the instructions here to install kaolin.

Download the tet-grid files (res128 & res256) to data/tets folder under the root directory. Alternatively, you may follow https://github.com/crawforddoran/quartet and data/tets/generate_tets.py to create the tet-grid files.

Generation

Install the following

• Pytorch3D
• ml_collections

To-dos

• [x] Code for reconstruction
• [ ] DeepFashion3D multiview image dataset for metallic surfaces
• [ ] Code for generative models
• [ ] Code for DeepFashion3D dataset preparation
• [ ] Evaluation code for generative models

Reconstruction

Datasets

DeepFashion3D mesh dataset

We provide ground-truth images (rendered under realistic environment light with Blender) for 9 instances in DeepFashion3D-v2 dataset. The download links for the raw meshes can be found in their repo.

non-metallic material: training data, test data

NeRF synthetic dataset

Download the NeRF synthetic dataset archive and unzip it into the data/ folder.

Hat dataset

Download link: https://drive.google.com/file/d/18UmT1NM5wJQ-ZM-rtUXJHXkDc-ba-xVk/view?usp=sharing

RGB images, segmentation masks and the corresponding camera poses are included. Alternatively, you may choose to 1) generate the camera poses with COLMAP and 2) create binary segmentation masks by yourself.

Training

DeepFashion3D-v2 instances

The mesh instances' IDs are [30, 92, 117, 133, 164, 320, 448, 522, 591]. To reconstruct the $INDEX-th mesh (0-8) in the list using tet-based G-Shell, run

  python train_gshelltet_deepfashion.py --config config/deepfashion_mc_256.json --index $INDEX --trainset_path $TRAINSET_PATH --testset_path $TESTSET_PATH --o $OUTPUT_PATH

For FlexiCubes + G-Shell, run

  python train_gflexicubes_deepfashion.py --config config/deepfashion_mc_80.json --index $INDEX --trainset_path $TRAINSET_PATH --testset_path $TESTSET_PATH --o $OUTPUT_PATH

Synthetic data

  python train_gshelltet_synthetic.py --config config/nerf_chair.json --o $OUTPUT_PATH

Hat data

  python train_gshelltet_polycam.py --config config/polycam_mc_128.json --trainset_path $TRAINSET_PATH --o $OUTPUT_PATH

  python train_gshelltet_polycam.py --config config/polycam_mc_128.json --trainset_path $TRAINSET_PATH --o $OUTPUT_PATH

On config files

You may consider modify the following, depending on your demand:

• gshell_grid: the G-Shell grid size. For tet-based G-Shell, please make sure the corresponding tet-grid file exists under data/tets (e.g., 256_tets.npz). Otherwise, follow https://github.com/crawforddoran/quartet and data/tets/generate_tets.py to generate the desired tet-grid file.
• n_samples: the number of MC samples for light rays per rasterized pixel. The higher the better (at a cost of memory and speed).
• batch_size: how many views sampled in each iteration.
• iteration: total number of iterations.
• kd_min, kd_max, etc: the min/max of the corresponding PBR material parameter.

Generation (To be done)

Citation

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

@article{liu2024gshell,
    title={Ghost on the Shell: An Expressive Representation of General 3D Shapes},
    author={Liu, Zhen and Feng, Yao and Xiu, Yuliang and Liu, Weiyang and Paull, Liam and Black, Michael J. and Sch{\"o}lkopf, Bernhard},
    booktitle={The Twelfth International Conference on Learning Representations},
    year={2024},
}

Acknowledgement

We sincerely thank the authors of Nvdiffrecmc, FlexiCubes and https://github.com/yang-song/score_sde_pytorch for sharing their codes. Our repo is adapted from MeshDiffusion.