Spherical Structure-from-Motion

Elfin Forest

Code for our papers:

Ventura, J., V. Larsson, F. Kahl. “Uncalibrated Structure from Motion on a Sphere”, International Conference on Computer Vision (ICCV), 2025.

Baker, L., S. Mills, S. Zollmann, and J. Ventura, “CasualStereo: Casual Capture of Stereo Panoramas with Spherical Structure-from-Motion”, IEEE Conference on Virtual Reality and 3D User Interfaces (VR), 2020.

Ventura, J., “Structure from Motion on a Sphere”, European Conference on Computer Vision (ECCV), 2016.

Dependencies

OpenCV 3+
Ceres solver
Eigen 3+
Polynomial
CUDA (for stereo panorama stitcher)

Building

mkdir build ;
cd build ;
cmake .. ;
make -j8 ;

The panorama stitcher app is disabled by default, since it requires CUDA which may not be available on all machines. To enable it:

mkdir build ;
cd build ;
cmake .. -DBUILD_STITCHER=TRUE ;
make -j8 ;

Usage

Calibrated Pipeline

If your images have radial distortion, first undistort them using:

undistort_images -intrinsics <path to intrinsics with distortion> -intrinsicsout <path for output intrinsics file> -video <path to video> -output <path ot folder for undistorted images> -rotate <rotation>

The “rotate” parameter is the number of clockwise 90 degree rotations to apply (useful for iPhone videos, for example).

To run the spherical structure-from-motion pipeline:

run_spherical_sfm -intrinsics <path to intrinsics> -video <path to video> -output <path to output>

The video path can be an image filename specifier such as %06d.png. The intrinsics file should contain the focal length, center x, and center y separated by spaces.

To make the stereo panoramas:

make_stereo_panorama -intrinsics <path to intrinsics> -video <path to video> -output <path to output>

Uncalibrated Pipeline

The uncalibrated pipeline supports estimating a single shared focal length.

run_spherical_sfm_uncalib \
    -images <path to video> \
    -output <path to output>

Alternately, you can use COLMAP to extract and match features, and then use our pipeline to perform the reconstruction.

PREFIX=<path to your dataset>

colmap feature_extractor \
    --database_path $PREFIX/database.db \
    --image_path $PREFIX/images \
    --ImageReader.camera_model=SIMPLE_PINHOLE \
    --ImageReader.single_camera=1

colmap exhaustive_matcher \
    --database_path $PREFIX/database.db

run_spherical_sfm_uncalib \
    -output $PREFIX \
    -images $PREFIX/images \
    -colmap \
    -sequential \
    -generalba

Capture tips

Hold the camera in an outstretched hand and turn in a roughly level and complete circle (close the loop). Don’t go too far over the loop closure point (no more than 30 degrees beyond the starting point).

Examples

You can view these example panoramas in a WebVR-compatible browser or headset.

PhoneSweep Dataset

Find the PhoneSweep Dataset from our ICCV 2025 paper hosted on OSF.

The dataset contains 13 outward spherical motion captures with pseudo-ground truth camera poses and intrinsics.