TexIR

This is official code for the paper accepted to CVPR 2023, "Multi-view Inverse Rendering for Large-scale Real-world Indoor Scenes".

Project Page | Paper

---

1. Installation

We encourage users to run our code in the Docker, which ensures the reproducibility of our implementation.

First, please follow nvdiffrec to build a basic docker image.

Second, starting an interactive docker container: docker run --runtime nvidia -it --rm --user root --gpus device=0 -v /your_data_folder:/data -v /code_folder:/code nvdiffrec:v1 /bin/bash

Finally, installing python packages in the container. pip install -r requirements.txt

2. Dataset

We release the real dataset and synthetic dataset proposed in the paper.

Please send an email to yodlee@mail.nwpu.edu.cn with some necessary information and we will send a download link to you.

# the necessary information of the request email
dataset: texir
name: xx
organization: xx
email for receiving link: xx

3. Running in our data

We also release our precomputed irradiance, both NIrF and IrT. Therefore, users could directly run the Material estimation with IrT. Users also could reproduce the NIrF and IrT via PIR.

3.1 Material estimation

For the synthetic dataset, please change the value of 'path_mesh_open3d' into the actual path in the container and run:

python3 trainer/exp_runner.py --conf configs/syn.conf --expname syn1 --trainstage MatSyn --gpu auto

For the real dataset, please change the value of 'path_mesh_open3d' into the actual path in the container and change the value of 'hdr_exposure'. Scene 1 and Scene 7 use exposure 3 and others use exposure 5. By the way, please change the threshold in the loss.py according to comments.

Finally, please run:

python3 trainer/exp_runner.py --conf configs/mat_hdrhouse.conf --expname scene1 --trainstage Mat --gpu auto

3.2 PIR (precomputed irradiance)

NIrF

For the real dataset, please change the value of 'path_mesh_open3d', similar to material estimation. Please run:

# NIrF
python3 trainer/exp_runner.py --conf configs/irrf_hdrhouse.conf --expname scene1 --trainstage IRRF --gpu auto

IrT

First, generating the IrT via ray tracing with 2048 samples.

There are some black regions in the IrT due to the error of UV mapping. Therefore, we do a simple padding to fill these regions in the texture space.

Finally, we denoise the IrT to reduce noise produced by monte carlo sampling.

Please run:

# 1. generate noised IrT, change hdr_exposure in hdrhouse_ir_texture.conf
python3 trainer/exp_runner.py --conf configs/hdrhouse_ir_texture.conf --expname scene1 --trainstage IrrT --gpu auto
# 2. padding IrT. change the path of tools/padding_texture.py and denoise the IrT via a denoiser, e.g. oidn, optix.
python3 tools/padding_texture.py

3.3 Evaluating previous methods

The sota multi-view object-centric neural rendering methods are InvRender, nvdiffrec and NeILF.

As described in the paper, we integrate their material optimization strategies with our lighting representation to handle large-scale indoor scenes.

Please run:

# note change exposure in .conf
# # real data
# invrender
python3 trainer/exp_runner.py --conf configs/mat_hdrhouse_invrender.conf --expname scene1 --trainstage Inv --gpu auto
# nvdiffrec
python3 trainer/exp_runner.py --conf configs/mat_hdrhouse_rec.conf --expname scene1 --trainstage RecMLP --gpu auto
# neilf
python3 trainer/exp_runner.py --conf configs/mat_hdrhouse_neilf.conf --expname scene1 --trainstage Neilf --gpu auto

# # synthetic data
# invrender
python3 trainer/exp_runner.py --conf configs/syn_invrender.conf --expname scene1 --trainstage InvSyn --gpu auto
# nvdiffrec
python3 trainer/exp_runner.py --conf configs/syn_rec.conf --expname scene1 --trainstage RecMLPSyn --gpu auto
# neilf
python3 trainer/exp_runner.py --conf configs/syn_neilf.conf --expname scene1 --trainstage NeilfSyn --gpu auto

4. Running in custom data

Our code is not support to run in other custom datasets, e.g. Replica. If you want to run our method with your data, please re-implement the interface or convert your data into our supported format, which consists of meshs, HDR images, UV-mapping, poses and segmentations.

Citation

We use the basic docker provided in nvdiffrec, please cite both our paper and their paper if you use our code.

@inproceedings{li2022texir,
    title={Multi-view Inverse Rendering for Large-scale Real-world Indoor Scenes},
    author={Li, Zhen and Wang, Lingli and Cheng, Mofang and Pan, Cihui and Yang, Jiaqi.},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.}, 
    year={2023}
}

@inproceedings{munkberg2021nvdiffrec,
    author = {Munkberg, Jacob and Hasselgren, Jon and Shen, Tianchang and Gao, Jun and Chen, Wenzheng and Evans, Alex and Mueller, Thomas and Fidler, Sanja},
    title = {Extracting Triangular 3D Models, Materials, and Lighting From Images},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.},
    year = {2022}
}