This repository contains the official implementation of 3R-GS, introduced in our paper:

🌐 Project Page: https://zsh523.github.io/3R-GS/

3R-GS: Best Practice in Optimizing Camera Poses Along with 3DGS
Zhisheng Huang, Peng Wang, Jingdong Zhang, Yuan Liu, Xin Li, Wenping Wang
arXiv:2504.04294

To set up the environment:

conda create --name 3rgs python=3.11 -y
conda activate 3rgs
conda install pytorch torchvision pytorch-cuda=12.1 -c pytorch -c nvidia -y
pip install -r requirements.txt

Download the original datasets:

• Tanks and Temples (TnT)
• Mip-NeRF 360
• DTU

We provide precomputed MASt3R-SfM results for each dataset:
📁 Google Drive

After downloading, extract each archive to the corresponding dataset root:

tar -xzf MipNeRF360.tar.gz --strip-components=1 -C YOUR_MIPNERF360_ROOT_PATH
tar -xzf TnT.tar.gz        --strip-components=1 -C YOUR_TNT_ROOT_PATH
tar -xzf DTU.tar.gz        --strip-components=1 -C YOUR_DTU_ROOT_PATH

Each scene directory will have the following structure:

your_dataset/
└── scene/
    ├── images/               # Original RGB images
    ├── sparse/               # Original GT annotations
    ├── mast3r/               # MASt3R-SfM outputs
    ├── images_train.txt      # Training split list
    ├── images_test.txt       # Testing split list
    ├── pose_gt_train.npy     # Ground truth train poses
    └── pose_gt_test.npy      # Ground truth test poses

Note: The MASt3R-SfM pipeline is not yet integrated into this codebase. For now, please refer to the MASt3R GitHub for details.

python src/trainer.py <mode> \
    --data_dir <INPUT_SCENE_PATH> \
    --data_factor <IMAGE_DOWNSAMPLE_RATE> \
    --result_dir <OUTPUT_PATH> \
    --pose_opt_type <pose_mode> \
    [--use_corres_epipolar_loss | --no-use_corres_epipolar_loss] \
    [--ckpt <CHECKPOINT_PATH>]

• <mode>: Pose optimization mode

  • default: Original 3DGS optimization
  • mcmc: MCMC-based 3DGS optimization

• --data_dir: Path to the input scene (e.g., ${TNT_ROOT}/Truck)

• --data_factor: Image downsampling factor (e.g., 1, 2, 4)

• --result_dir: Output directory for saving results

• --pose_opt_type: Pose optimization method

  • sfm: Optimize camera poses directly
  • mlp: Use MLP-based global pose refinement

• --use_corres_epipolar_loss or --no-use_corres_epipolar_loss: Whether to apply global epipolar loss

• --ckpt: (Optional) Path to a checkpoint for evaluation

• Naive joint optimization (baseline):

  bash scripts/3dgs_train.sh

• Our method (3R-GS):

  bash scripts/3rgs_train.sh

• 3D Gaussian Splatting code is based on gsplat
• Evaluation scripts adapted from MonoGS

If you find our project helpful, please consider citing:

@misc{huang20253rgsbestpracticeoptimizing,
  title     = {3R-GS: Best Practice in Optimizing Camera Poses Along with 3DGS},
  author    = {Zhisheng Huang and Peng Wang and Jingdong Zhang and Yuan Liu and Xin Li and Wenping Wang},
  year      = {2025},
  eprint    = {2504.04294},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url       = {https://arxiv.org/abs/2504.04294}
}