PDF

This repository contains the code for PBDL 2024 (CVPR Workshop) paper: Generating Material-Aware 3D Models from Sparse Views

MatSparse3D introduces a novel approach to generate material-aware 3D models from sparse-view images using generative models and efficient pre-integrated rendering. The output of our method is a relightable model that independently models geometry, material, and lighting, enabling downstream tasks to manipulate these components separately.

create envrionment using mamba or conda. And install additional packages using pip.

mamba env create --file environment.yml
pip install -r requirements_git+.txt 

Our method use pretrained Zero123-XL model, the pretrained Zero123-XL weights are required to be download and save in load/zero123:

mkdir load/zero123
cd load/zero123
wget https://zero123.cs.columbia.edu/assets/zero123-xl.ckpt

• DTU-MVS: Download the DTU sample set from DTU-MVS. The depth and masking can be estimated by running preprocessing.
• sparse-relight: We provide our generated sparse-relight scenes on Google Drive. The provided data has estimated depth without the need for preoprocessing

 DATA/
 ├── DTU-MVS/
 │   └── SampleSet
 ├── sparse-relight/
     ├── light-probes/
     ├── mesh/
     └── synthesis-images/
          ├── cartooncar/
          ├── gramophone/
          └── ......

Omnidata is used for depth and normal prediction. The following ckpts are hardcoded in preprocess_image.py and required to be downloaded:

mkdir load/omnidata
cd load/omnidata
# dowload omnidata_dpt_depth_v2.ckpt
gdown '1Jrh-bRnJEjyMCS7f-WsaFlccfPjJPPHI&confirm=t' 
## optionally download omnidata_dpt_normal_v2.ckpt for normal prediction
# gdown '1wNxVO4vVbDEMEpnAi_jwQObf2MFodcBR&confirm=t' 

Preprocessing data, with optional --use_normal to estimate the normal if needed.

# e.g. for sparse-relight data
python preprocess.py --scene_dir DATA/sparse-relight/cartooncar
# e.g. for DTU data
python preprocess_DTU.py --scene_dir DATA/DTU-MVS/SampleSet/MVS-Data/Rectified/scan56

Start training by running launch.py with --train flag.

• Number of train views can be specified by data.tran_views.
• Config file for DTU and sparse-relightdataset are available

For more settings, please refer to the config file.

# to train the proposed MatSparse3D model on Sparse-Relight dataset
python launch.py --config configs/matsparse3d_sparserelight.yaml --train --gpu 0 data.train_views=5
# to train the proposed MatSparse3D model on DTU dataset
python launch.py --config configs/matsparse3d_DTU.yaml --train --gpu 0 data.train_views=5
# to train the Zero123-n model
python launch.py --config configs/zero123n_sparserelight.yaml --train --gpu 0 data.train_views=5
# to train the nvdiffrec-n model
python launch.py --config configs/nvdiffrec_sparserelight.yaml --train --gpu 0 data.train_views=5

use --validate to render interpolated views, an examle is provided in script_test.sh

# beware to specify the exp_folder
bash script_validate.sh 

use --test to evaluate on testing views. For sparse-relight dataset, relighting result is also reported. An examle is provided in script_test.sh

# beware to specify the exp_folder
bash script_test.sh 

use --export to export geometry from trained model, an examle is provided in script_export.sh

# beware to specify the exp_folder
bash script_export.sh 

Matsparse3D is built on the following open-source projects:

• Threestudio A unified framework for 3D content creation
• stable-dreamfusion for Zero-1-to-3 implementation
• NeuSPIR Learning Relightable Neural Surface using Pre-Integrated Rendering

If MatSparse3D is relevant to your project, please cite our associated paper:

@InProceedings{mao2023matsparse3d,
  author={Mao, Shi and Wu, Chenming and Yi, Ran and Shen, Zhelun and Zhang, Liangjun and Heidrich, Wolfgang},
  title={Generating Material-Aware 3D Models from Sparse Views},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  year={2024},
  month={June},
}