UrbanIR: Large-Scale Urban Scene Inverse Rendering from a Single Video

3DV 2025

Project Page | Paper | Data | Checkpoints

Chih-Hao Lin¹, Bohan Liu¹, Yi-Ting Chen², Kuan-Sheng Chen¹,

David Forsyth¹, Jia-Bin Huang², Anand Bhattad¹, Shenlong Wang¹

¹University of Illinois at Urbana-Champaign, ²University of Maryland, College Park

🔦 Prerequisites

The code has been tested on: - OS: Ubuntu 22.04.4 LTS - GPU: NVIDIA GeForce RTX 4090, NVIDIA RTX A6000 - Driver Version: 535, 545 - CUDA Version: 12.2, 12.3 - nvcc: 11.7

🔦 Installation

• Create Conda environment:

conda create -n urbanir -y python=3.9
conda activate urbanir

• Install python packages:

pip install -r requirements.txt

• Install pytorch_scatter:

pip install torch-scatter -f https://data.pyg.org/whl/torch-2.0.1+cu117.html

• Install tiny-cuda-nn:

git clone --recursive https://github.com/NVlabs/tiny-cuda-nn.git

Then use your favorite editor to edit tiny-cuda-nn/include/tiny-cuda-nn/common.h and set TCNN_HALF_PRECISION to 0 (see NVlabs/tiny-cuda-nn#51 for details)

cd tiny-cuda-nn/bindings/torch
python setup.py install

• Compile CUDA extension of this project

pip install models/csrc/

🔦 Dataset and Checkpoints

• Please download datasets and put under data/ folder.
• Please download checkpoints and put under ckpts folder.
• Currently the data and checkpoints of Kitti360 and Waymo Open Dataset are available.

🔦 Training

• The training process is tracked and visualized with wandb, you could set up by $ wandb login
• The training script examples are in scripts/train.sh, which are in the following format:

python train.py --config [path_to_config]

• The checkpoint is saved to ckpts/[dataset]/[exp_name]/*.ckpt
• The validation images are saved to results/[dataset]/[exp_name]/val

🔦 Rendering

• The rendering script examples are in scripts/render.sh, which are in the following format:

python render.py --config [path_to_config]

• The rendered images are saved to results/[dataset]/[exp_name]/frames, and videos are saved to results/[dataset]/[exp_name]/*.mp4
• If the training is not complete and .../last_slim.ckpt is not available, you can either specify path with --ckpt_load [path_to_ckpt] or convert with utility/slim_ckpt.py.

🔦 Relighting

• The relighting script examples are in scripts/relight.sh, which are in the following format:

python render.py --config [path_to_config] \
     --light [path_to_light_config] --relight [effect_name]

• The rendered images and videos are saved to results/[dataset]/[exp_name]/[effect_name]

🔦 Configuration

• All the parameters are listed in the opt.py, and can be added after training/rendering/relighting script with --param_name value.
• The scene-specific parameters are listed in configs/[dataset]/[scene].txt.
• The lighting parameters are listed in configs/light/[scene].txt, and different relighting effects can be produced by changing the configuration.

🔦 Customized Data

• The camera poses are estimated with NeRFstudio pipeline (transforms.json).
• The depth is estimated with MiDaS.
• The normal is estimated with OmniData, and authors' fork can handle image resolution in a more flexible manner.
• The shadow mask is estimated with MTMT, and authors' fork provides example script.
• The semantic map is estimated with mmsegmentation, and you could put mmseg/run.py in the root folder of mmsegmentation and run.

🔦 Citation

If you find this paper and repository useful for your research, please consider citing:

@article{lin2023urbanir,
  title={Urbanir: Large-scale urban scene inverse rendering from a single video},
  author={Lin, Zhi-Hao and Liu, Bohan and Chen, Yi-Ting and Forsyth, David and Huang, Jia-Bin and Bhattad, Anand and Wang, Shenlong},
  journal={arXiv preprint arXiv:2306.09349},
  year={2023}
}