g2o

g2o stands for General Graph Optimization. It is an open-source C++ framework designed to solve graph-based nonlinear least-squares problems, with a focus on pose graph optimization in robotics and computer vision. The library is commonly used to formulate problems as a graph: vertices correspond to variables such as poses or landmarks, and edges encode constraints derived from sensor measurements.

Its architecture is modular. Vertex and Edge base classes allow users to define problem-specific variables and residuals. Edges carry residual functions and information matrices and support different Lie groups, including SE(2) and SE(3) for pose graphs. The framework implements optimization algorithms such as Gauss-Newton and Levenberg–Marquardt, and exposes a flexible solver stack with interchangeable linear solvers and block solvers. It relies on sparse linear algebra and can use backends such as Cholmod or CSparse through optional dependencies. Robust kernels (Huber, Cauchy, and others) can be attached to edges to reduce sensitivity to outliers.

g2o was introduced by Grisetti, Kummerle, Stachniss, and Burgard in 2010 as a general framework to simplify and accelerate graph-based optimization. It has become a standard tool in SLAM, visual odometry, and related computer vision research, and is used in both academic projects and industry prototypes. The library is open-source and widely distributed, with ongoing updates and community contributions via its repository. Users typically construct a graph by adding vertices for states and edges for measurements, select a solver and robust options, and solve to obtain the optimized estimates.