/pyceres

Factor graphs with Ceres in Python

Primary LanguageC++

This is a test(fork) version of pyceres that adds Sim3 Cost function

This repository provides minimal Python bindings for the Ceres Solver and the implementation of factor graphs for bundle adjustment and pose graph optimization.

Additional Dependencies besides original dependencies for pyceres

  1. Sophus
  2. (optional, for Sophus) fmt ,if you have problems with fmt when building pyceres, try downgrading fmt to 8.0.0 or build Sophus without fmt.

Installation

Before installation, make sure you have Sophus and fmt.

  1. Clone the repository and its submodule by running:
git clone --recursive git@github.com:kaikai23/pyceres.git
cd pyceres

  1. Install COLMAP.

  2. Build the package:

pip install -e .

Factor graph optimization

For now we support the following cost functions, defined in _pyceres/factors/:

  • camera reprojection error (with fixed or variable pose)
  • rig reprojection error (with fixed or variable rig extrinsics)
  • relative pose prior
  • absolute pose prior

All factors support basic observation covariances. Reprojection error costs rely on camera models defined in COLMAP. Absolute poses are represented as quaternions and are expressed in the sensor frame, so are pose residuals, which use the right-hand convention as in the GTSAM library.

Examples

See the Jupyter notebooks in examples/.

TODO

  • Define a clean interface for covariances, like in GTSAM
  • Add bindings for Ceres covariance estimation
  • Use proper objects for poses, e.g. from Sophus
  • Proper benchmark against GTSAM

Credits

The core bindings were written by Nikolaus Mitchell for ceres_python_bindings and later adapted by Philipp Lindenberger for pixel-perfect-sfm.