[ICRA 2026] PROFusion: Robust and Accurate Dense Reconstruction via Camera Pose Regression and Optimization

Siyan Dong · Zijun Wang · Lulu Cai · Yi Ma · Yanchao Yang

Paper | Video | Poster

A simple yet effective system for real-time camera tracking and dense scene reconstruction, providing both robustness against unstable camera motions and accurate reconstruction results.

Installation

1. Clone PROFusion

git clone https://github.com/siyandong/PROFusion.git
cd PROFusion

2. Prepare environment

conda create -n profusion python=3.11 cmake=3.14 -y
conda activate profusion

# install torch according to your cuda version, e.g., 12.4
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu124

# install packages required by the pose regression module
pip install -r requirements.txt

# install packages required by the optimization and fusion module
conda install -y -c conda-forge gcc_linux-64=11.4.0 gxx_linux-64=11.4.0 pybind11 

Install Eigen and OpenCV with CUDA required by the optimization and fusion module, following this link

3. Build PROFusion: Set sm_xx and compute_xx in L30 in CMakeLists.txt according to your architecture and run the following script

bash build.sh

4. Optional: If you cannot load the pretrained model automatically, manually download and format the directory as shown in L11 in pose_regression/run.py

5. Optional: Acceleration

# compile cuda kernels for RoPE
cd pose_regression/modules/pos_embed/curope/
python setup.py build_ext --inplace
cd ../../../../

Run demos

Download example data from Google Drive and format directories as:

data/
├── femtobolt/
├── ├── p207/
├── ├── ├── color/
├── ├── ├── depth/
├── ├── p302_sparse/
├── fastcamo/example/
└── eth3d/cs1/

Otherwise, you can edit these paths in configs/data_config/xx.yaml

FemtoBolt

python run_profusion.py configs/data_config/femtobolt.yaml

FastCaMo-Synth (noise)

python run_profusion.py configs/data_config/fastcamo.yaml

ETH3D

python run_profusion.py configs/data_config/eth3d.yaml

Citation

If you find our work helpful in your research, please consider citing:

@article{dong2025profusion,
  title={PROFusion: Robust and Accurate Dense Reconstruction via Camera Pose Regression and Optimization},
  author={Dong, Siyan and Wang, Zijun and Cai, Lulu and Ma, Yi and Yang, Yanchao},
  journal={arXiv preprint arXiv:2509.24236},
  year={2025}
}

Acknowledgments

Our implementation is based on: DUSt3R, SLAM3R, Reloc3r, ROSEFusion, and many other inspiring works in the community.

This is an open-source version of PROFusion with a reorganized architecture and rewritten functions for easier use. While it may not reproduce the original results exactly, the output is nearly identical. The code has been tested on several devices with CUDA 12.4, including 4090, A800, and H800.