CVD-SfM

This repository accompanies the paper CVD-SfM: A Cross-View Deep Front-end Structure-from-Motion System for Sparse Localization in Multi-Altitude Scenes (IROS 2025). If you find this work useful please cite out paper.

@misc{li2025cvdsfmcrossviewdeepfrontend,
      title={CVD-SfM: A Cross-View Deep Front-end Structure-from-Motion System for Sparse Localization in Multi-Altitude Scenes}, 
      author={Yaxuan Li and Yewei Huang and Bijay Gaudel and Hamidreza Jafarnejadsani and Brendan Englot},
      year={2025},
      eprint={2508.01936},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.01936}, 
}

We developed a Cross-View Deep Front-end Structure-from-Motion System for Sparse Localization in Multi-Altitude Scenes. The performance of our approach is shown in the video.

We also create 2 multi-altitude datasets with ground truth GPS for two different sites. You can download datasets on hugging face: https://huggingface.co/datasets/yaxlee/Stevens-Sky2Ground.

If you find this repository useful, please cite our paper:

Benchmark

We visualize pose estimation results on three different datasets. Yellow Points represent estimated poses; cyan points represent ground truth poses.

Reconstructed Model review on SIT Campus dataset. Red frames represent aerial images, blue frames represent ground images, black frame on the top represent reference satellite image.

Installation

Our implementaion is on Ubuntu20.04, python=3.11.5, torch=2.0.0, torchvision=0.15.1.

COLMAP

git clone https://github.com/RobustFieldAutonomyLab/CVD-SfM.git
cd CVD-SfM/colmap
mkdir build && cd build
cmake .. \
  -DCMAKE_BUILD_TYPE=Release \
make -j$(nproc)
sudo make install

Dependence

pip install -r requirements.txt

Run Instruction

python run_cvd_sfm.py

Remeber to change the root dir to your own path and make sure /images and /sat dir are included in root dir.

└── Root Dir
      ├── images
               ├── image1
               ├── image2
               ├── ..
      ├──sat
               ├── satellite image

Custom-Gathered Dataset: Stevens-Sky2Ground

We collect two multi-altitude datasets with ground truth GPS for two different sites. Each contains aerial imagery collected by UAV and ground imagery collected by handheld device. One high-resolution satellite imagery from Google Earth Pro is also included for each site. Ground-level GPS is achieved by RTK GNSS using EMLID Reach RS+ receivers.

Stevens Institute of Technology Campus

                 

This dataset contains 179 aerial images, 186 ground images and 1 satellite image.

Raritan Bay Waterfront Park

                 

This dataset contains 174 aerial images, 139 ground images and 1 satellite image.

Datasets Structure:

└── SIT campus
       ├── images
               ├── aerial_image1.png
               ├── aerial_image2.png
               ├── ..
               ├── ground_image1.png
               ├── ground_image2.png
               ├── ...
      ├── gps
               ├── aerial_image1.json
               ├── aerial_image2.json
               ├── ...
               ├── ground_image1.json
               ├── ground_image2.json
               ├── ...
      ├── satellite_sit.jpg
      ├── trajectory.jpg
└── Raritan bay
       ├── ...

Our dataset is available on hugging face: https://huggingface.co/datasets/yaxlee/Stevens-Sky2Ground