ceres-solver

ceres-solver-rs

Rust bindings for Ceres Solver

cargo add ceres-solver --features=source

Ceres Solver is a C++ library for large optimization problems. It can be used to solve Non-linear Least Squares problems with constraints and general optimization problems. Here we provide a Rust binding for this library.

The bindings require Ceres Solver version 2.2, but the bindings may work with older or new versions. Please submit an issue if you need a support of other versions.

This project consists of three crates:

• ceres-solver is a safe Rust bindings
• ceres-solver-sys is an unsafe Rust bindings written with the usage of cxx
• ceres-solver-src is an optional no-code crate to build and distribute a minimal static Ceres Solver library

To build Ceres Solver statically and link it to your project, use source Cargo feature, which would add ceres-solver-src dependency into your project.

Status of the binding support

Current implementation of the binding is not complete. The following list shows the status of the binding support:

• Non-linear Least squares
  • Problem - basic class for NLLS, supports adding residual blocks, setting boundary conditions, marking parameter blocks to be constant/variable, and solving the problem
  • CostFunction - user provides both residual and Jacobian
  • SizedCostFunction - same but with the residual vector shape is known at compile time
  • AutoDiffCostFunction - user provides residual and Jacobian is computed by automatic differentiation
  • DynamicAutoDiffCostFunction - same but with the residual vector shape is unknown at compile time
  • NumericDiffCostFunction - user provides residual and Jacobian is computed by numerical differentiation
  • CostFunctionToFunctor and DynamicCostFunctionToFunctor - adapter to use CostFunction as a mix of all other cost functions
  • ConditionedCostFunction - adapter to use CostFunction with different conditioning
  • GradientChecker - helper class to check the correctness of the Jacobian
  • NormalPrior - changes a cost function to use a covariance matrix instead of a simple scalar product
  • LossFunction - a function applied to the squared norm of the residual vector, both custom and Ceres stack loss functions are supported
  • Manifold, AutoDiffManifold
  • EvaluationCallback
• Solver - Solver class itself is not implemented, but the following nested classes are supported:
  • Solver::Options
    • Minimizer options
    • Line search options
    • Trust region options
    • Linear solver options
    • Preconditioner options
    • Sparse and dense linear algebra library selection
    • Setting of the number of threads
    • Bundle adjustment options
    • Logging options
    • Validation of the options
    • Callbacks
  • Solver::Summary
    • Brief and full reports
    • Cost function evaluation statistics
    • Time statistics
• Jets
• Covariance estimation
• General unconstrained minimization

Please don't hesitate to create an issue to request prioritization of any functionality you need.