EVolSplat: Efficient Volumetric Splatting for Real-Time Urban View Synthesis (CVPR 2025)

Sheng Miao, Jiaxin Huang, Dongfeng Bai, Xu Yan, Hongyu Zhou, Yue Wang, Bingbing Liu, Andreas Geiger and Yiyi Liao

Our project page can be seen here.

📖 Datasets

We evaluate our model on KITTI-360 and Waymo. Here we show the structure of a test dataset as follow, similar to the EDUS. We provide the one example data for inference on KITTI-360, which can be found in huggingface here.

The dataset should have a structure as follows:

├── $PATH_TO_YOUR_DATASET
    ├── $SCENE_0
        ├── depth
        ├── pointcloud/*.ply
        ├── *.png
        ...
        ├── transfroms.json
    ...
    ├── SCENE_N
        ├── depth
        ├── pointcloud/*.ply
        ├── *.png
        ...
        ├── transfroms.json

🏠 Installation

Our EVolSplat is built on nerfstudio. You can follow the nerfstudio webpage to install our code.

Create environment

We recommend using conda to create a new environment and you can find the detailed environment file in environment.yml.

conda create --name EVolSplat -y python=3.8
conda activate EVolSplat
pip install --upgrade pip

Dependencies

Install PyTorch

Install PyTorch with CUDA (this repo has been tested with CUDA 11.8 with torch 2.1.2 and CUDA 11.7 with torch 2.0.1).

For CUDA 11.8 with torch 2.1.2:

pip install torch==2.1.2+cu118 torchvision==0.16.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118
conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit

For CUDA 11.7 with torch 2.0.1:

pip install torch==2.0.1+cu117 torchvision==0.15.2+cu117 --extra-index-url https://download.pytorch.org/whl/cu117
conda install -c "nvidia/label/cuda-11.7.1" cuda-toolkit

Install TorchSparse

Install the spaseCNN library, we recommend the version 2.1.0. We can find the installation instruction in the torchsparse repository.

conda install -c conda-forge sparsehash
sudo apt-get install libsparsehash-dev  
git clone --recursive https://github.com/mit-han-lab/torchsparse
python setup.py install

Install gsplat

Install gsplat from source code. We recommend the version >= 1.0.0.

pip install ninja==1.11.1.1
pip install git+https://github.com/nerfstudio-project/gsplat.git

If you meet some problems, you can check the dependencies from environment.yml.

Install EVolSplat

Install EVolSplat form source code

git clone https://github.com/XDimLab/EVolSplat.git
cd EVolSplat
pip install --upgrade pip setuptools
pip install -e .

📈 Evaluation & Checkpoint

We provide the pretrained model trained on KITTI-360, you can download the pre-trained models and example data from huggingface for a quick start.

Place the downloaded checkpoints in checkpoints folder in order to test it later.

Feed-forward Inference

Replace $PATH_TO_YOUR_DATASET$ with your data path.

python nerfstudio/scripts/infer_zeroshot.py evolsplat \
  --load_dir checkpoints/ \
  --pipeline.model.freeze_volume=True \
  zeronpt-data \
  --data $PATH_TO_YOUR_DATASET$ \
  --kitti=True 

📋 Citation

If our work is useful for your research, please give me a star and consider citing:

@inproceedings{miao2025efficient,
  title={EVolSplat: Efficient Volumetric Splatting for Real-Time Urban View Synthesis},
  author={Miao, Sheng and Huang, Jiaxin and Bai, Dongfeng and Yan, Xu and Zhou, Hongyu and Wang, Yue and Liu, Bingbing and Geiger, Andreas and Liao, Yiyi},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
}

✨ Acknowledgement

• This project is based on nerfstudio and gsplat.
• Some codes are brought from IBRNet and SparseNeus.