GlobalSfMpy

This repo is the implementation of the paper "Revisiting Rotation Averaging: Uncertainties and Robust Losses".

If you find our code or paper useful, please cite

@inproceedings{zhang2023revisiting,
  title={Revisiting Rotation Averaging: Uncertainties and Robust Losses},
  author={Zhang, Ganlin and Larsson, Viktor and Barath, Daniel},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={17215--17224},
  year={2023}
}

Tested Environment

Ubuntu 22.04

Dependency

• TheiaSfM (Recommand version 0.7.0)
• Ceres (Recommand version 1.14.0)
• pybind11 (Recommand version 2.9.2)
• OpenCV
• (Optional) COLMAP (Recommand version 3.6)

Ceres

More details can be found on the Ceres official website

# install dependencies of Ceres Solver
sudo apt-get install cmake libgoogle-glog-dev libgflags-dev libatlas-base-dev libeigen3-dev libsuitesparse-dev

# download and install Ceres Solver
wget https://github.com/ceres-solver/ceres-solver/archive/refs/tags/1.14.0.tar.gz
tar zxf 1.14.0.tar.gz
mkdir ceres-bin
cd ceres-bin
cmake ../ceres-solver-1.14.0
make -j3
make install

TheiaSfM

More details can be found on the TheiaSfM official website

# install dependencies of Theia
sudo apt-get install libopenimageio-dev librocksdb-dev libatlas-base-dev rapidjson-dev libgtest-dev libyaml-cpp-dev
cd thirdparty/TheiaSfM
mkdir build && cd build
cmake ..
make -j4
sudo make install

PyBind11

git clone git@github.com:pybind/pybind11.git
cd pybind11 && mkdir build && cd build
cmake .. && make -j4
sudo make install

OpenCV

More details can be found on the OpenCV official website

# Install minimal prerequisites (Ubuntu 18.04 as reference)
sudo apt update && sudo apt install -y cmake g++ wget unzip
# Download and unpack sources
wget -O opencv.zip https://github.com/opencv/opencv/archive/4.x.zip
wget -O opencv_contrib.zip https://github.com/opencv/opencv_contrib/archive/4.x.zip
unzip opencv.zip
unzip opencv_contrib.zip
# Create build directory and switch into it
mkdir -p build && cd build
# Configure
cmake -DOPENCV_EXTRA_MODULES_PATH=../opencv_contrib-4.x/modules ../opencv-4.x
# Build
cmake --build .

Install

mkdir build && cd build
cmake ..
make

Demo

1DSfM Madrid_Metropolis

cd script
python sfm_pipeline.py ../flags_1dsfm.yaml

ETH3D terrace

First, use COLMAP extract the feature points and two-view matches. When extracting the feature, please use the PINHOLE camera model. Here a bash script is provided to run COLMAP for extracting the two-view information:

Download the datasets from here and put the images inside the datasets/terrace/images folder, i.e.

.
└── datasets
    └── terrace
        └── images
            └── *.JPG

First change the colmap executable path in the script, then

cd script
bash colmapFeatureMatching.sh ../datasets/terrace

After running COLMAP, the sturcture of the dataset folder should be as following:

.
└── datasets
    └── terrace
        ├── cameras.txt
        ├── colmap
        │   ├── 0
        │   │   ├── cameras.bin
        │   │   ├── images.bin
        │   │   ├── points3D.bin
        │   │   └── project.ini
        │   └── database.db
        ├── covariance_rot.txt
        ├── images
        │   └── *.JPG
        ├── images.txt
        ├── points3D.txt
        └── two_views.txt

Then, use GobalSfMpy to reconstruct the scene. The following commands are run inside script folder.

python read_colmap_database.py --dataset_path ../datasets/terrace
python get_covariance_from_colmap.py --dataset_path ../datasets/terrace
python sfm_with_colmap_feature.py --dataset_path ../datasets/terrace

The reconstruction is stored in output folder. The format of the result is the same as what it is in TheiaSfM. The Theia application view_reconstruction can be used to visualize the result.

./view_reconstruction --reconstruction <RESULT_FILE>

E.g. For 1DSfM Madrid_Metropolis

./thirdparty/TheiaSfM/build/bin/view_reconstruction --reconstruction output/Madrid_Metropolis

Example visualization of the output of 1DSfM Madrid_Metropolis:

Example visualization of the output of ETH3D terrace