COLMAP-Free 3D Gaussian Splatting

Yang Fu · Sifei Liu · Amey Kulkarni · Jan Kautz
Alexei A. Efros · Xiaolong Wang

Paper | Video | Project Page

Installation

(Recommended)

The codes have been tested on python 3.10, CUDA>=11.6. The simplest way to install all dependences is to use anaconda and pip in the following steps:

conda create -n cf3dgs python=3.10
conda activate cf3dgs
conda install conda-forge::cudatoolkit-dev=11.7.0
conda install pytorch==2.0.0 torchvision==0.15.0 pytorch-cuda=11.7 -c pytorch -c nvidia
git clone --recursive git@github.com:NVlabs/CF-3DGS.git
pip install -r requirements.txt

Dataset Preparsion

DATAROOT is ./data by default. Please first make data folder by mkdir data.

Tanks and Temples

Download the data preprocessed by Nope-NeRF as below, and the data is saved into the ./data/Tanks folder.

wget https://www.robots.ox.ac.uk/~wenjing/Tanks.zip

CO3D

Download our preprocessed data, and put it saved into the ./data/co3d folder.

Run

Training

python run_cf3dgs.py -s data/Tanks/Francis \ # change the scene path
                     --mode train \
                     --data_type tanks

Evaluation

# pose estimation
python run_cf3dgs.py --source data/Tanks/Francis \
                     --mode eval_pose \
                     --data_type tanks \
                     --model_path ${CKPT_PATH} 
# by default the checkpoint should be store in "./output/progressive/Tanks_Francis/chkpnt/ep00_init.pth"
# novel view synthesis
python run_cf3dgs.py --source data/Tanks/Francis \
                     --mode eval_nvs \
                     --data_type tanks \
                     --model_path ${CKPT_PATH} 

We release some of the novel view synthesis results (gdrive) for comparison with future works.

Run on your own video

• To run CF-3DGS on your own video, you need to first convert your video to frames and save them to ./data/$CUSTOM_DATA/images/

• Camera intrincics can be obtained by running COLMAP (check details in convert.py). Otherwise, we provide a heuristic camera setting which should work for most landscope videos.

• Run the following commands:

python run_cf3dgs.py -s ./data/$CUSTOM_DATA/ \ # change to your data path
                     --mode train \
                     --data_type custom

Acknowledgement

Our render is built upon 3DGS. The data processing and visualization codes are partially borrowed from Nope-NeRF. We thank all the authors for their great repos.

Citation

If you find this code helpful, please cite:

@InProceedings{Fu_2024_CVPR,
    author    = {Fu, Yang and Liu, Sifei and Kulkarni, Amey and Kautz, Jan and Efros, Alexei A. and Wang, Xiaolong},
    title     = {COLMAP-Free 3D Gaussian Splatting},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {20796-20805}
}