PanoGRF

This is the official code release for NeurIPS2023 paper, PanoGRF: Generalizable Spherical Radiance Fields for Wide-baseline Panoramas

Project Page|Arxiv

Update:

2024.2.2 upload the preprocess files for Replica and Residential. See the README files for replica and Residential

Citation

If you find this repo useful, please give me a star and cite this paper:

@article{chen2023panogrf,
  title={PanoGRF: Generalizable Spherical Radiance Fields for Wide-baseline Panoramas},
  author={Chen, Zheng and Cao, Yan-Pei and Guo, Yuan-Chen and Wang, Chen and Shan, Ying and Zhang, Song-Hai},
  journal={arXiv preprint arXiv:2306.01531},
  year={2023}
}

Environment

Refer to installation guidance

Dataset

Download

We download Matterport3D following SynSin.

Please fill and sign the Terms of Use agreement form and send it to [email protected] to request access to the dataset.

The offical download script (download_mp.py) will be given in the reply email after your request is approved.

The full MP3D dataset for use with Habitat can be downloaded using the official Matterport3D download script as follows: python download_mp.py --task habitat -o path/to/download/. Note that this download script requires python 2.7 to run.

Dataset Path Configuration

You should change the name of all the saving directories in the config files according to your directories.

You should also revise the data directory opts.scenes_dir according to your download directory which stored mp3d in the following files:

data_readers/habitat_data_neuray_ft.py
data_readers/habitat_data_neuray.py

The data for opts.scene_dir is organized as:

<opts.scene_dir>
|-- mp3d 
    |-- 1LXtFkjw3qL # scene_name
            |-- 1LXtFkjw3qL_semantics.ply
            |-- 1LXtFkjw3qL.glb
            |-- 1LXtFkjw3qL.house
            |-- 1LXtFkjw3qL.navmesh            
    |-- 1pXnuDYAj8r
            |-- ...
    |-- ...

pretrained model

The pretrained models of 360-degree Monocular Net, 360-degree MVSNet and general renderer(two-views trained under 1.0m camera baseline) can be found in GoogleDrive

Depth Training

Monocular depth finetuning:

Download the pretrained model on Matterport3D from UniFuse and put it under the directory load_weights_dir in config file configs/train/depth/m3d_mono.yaml.

Training the monocular depth network as follows:

bash train_scripts/train_depth/train_monodepth.sh

Multi-view stereo network training with monocular depth prior:

set DNET_ckpt in configs/train/depth/m3d_mono.yaml to the path of monocular depth model in last step.

bash train_scripts/train_depth/train_mvs.sh

General renderer training:

The speed of rendering training data (textured mesh) online with habitat is tolerable in depth training. But it is quite slow for training general renderer. So I used lmdb to pre-rendering and save the data here.

Data preprocessing using lmdb:

# preprocessing training data:
python lmdb_rw_render/lmdb_write_render.py --cfg configs/data/train_data_render_512x1024.yaml
# please preprocess val/test data similarly.

This step takes too much storage space. If the saved data is too large for you, try to reduce total_cnt in data-preprocessing config file and revise it in the training config files correspondingly.

bash train_scripts/gen_hr_1.0/gen_mono_stereo_uniform_512x1024.sh

render & eval with pretrained models (GoogleDrive):

In configuration file configs/train/gen_hr_1.0/neuray_gen_cv_erp_mono_stereo_uniform_512x1024.yaml, you need to

• ensure save_dir is the precessed data directory (lmdb format)

• revise DNET_ckpt -> pretrained monocular depth model path habitat_monodepth/checkpoint_100000.pt

• revise mvsnet_pretrained_path -> pretrained MVS depth model path habitat_mvs/checkpoint_100000.pt

• run mkdir -p data/neuray_gen_erp_1.0_mono_stereo_uniform_512x1024, put pretrained renderer model general_renderer/model.pth into data/neuray_gen_erp_1.0_mono_stereo_uniform_512x1024

Then run the following command to get renderer results.

bash render_scripts/gen_hr_1.0/gen_eval_m3d.sh 

Todo List

• multi-view training
• fine-tune training
• Dataset: Replica
• Dataset: Residential
• clean up unnessary codes
• ......

Acknowledgements

Within this repository, I have utilized code and datasets from various sources. I express my gratitude to all the authors who have generously shared their valuable resources, enabling me to build upon their work:

• SOMSI
• OmniSyn
• MaGNet
• UniFuse
• SynSin
• Matterport3D
• Replica
• Residential