๐ฆ Apex Solver
A high-performance Rust-based nonlinear least squares optimization library designed for computer vision applications including bundle adjustment, SLAM, and pose graph optimization. Built with focus on zero-cost abstractions, memory safety, and mathematical correctness.
Apex Solver is a comprehensive optimization library that bridges the gap between theoretical robotics and practical implementation. It provides manifold-aware optimization for Lie groups commonly used in computer vision, multiple optimization algorithms with unified interfaces, flexible linear algebra backends supporting both sparse Cholesky and QR decompositions, and industry-standard file format support for seamless integration with existing workflows.
โ ๏ธ Upgrading to 1.4.0 โ breaking API changes
1.4.0 changes the public API. Code written against 1.3.0 will not compile until you make
the edits below. Full detail in the changelog.
1. Problem uses handles instead of string names. add_variable returns a VarKey;
add_residual_block takes &[VarKey] and returns a FactorKey (previously &[&str] and
usize). Keep the returned key and pass it where you used to pass a name:
// 1.3.0
problem.add_variable;
problem.add_residual_block;
// 1.4.0
let k0 = problem.add_variable;
let k1 = problem.add_variable;
problem.add_residual_block;
If you need to look variables up later, keep your own HashMap<YourId, VarKey> โ the
Quick Start below shows the pattern.
2. Factor::get_dimension is renamed to Factor::residual_dim. Custom factor
implementations must rename the method; there is no default implementation.
// 1.3.0 // 1.4.0
3. OptimizationStatus gained a StalledNoProgress variant. Exhaustive match
expressions need a new arm. Treat it as a successful termination โ it means the solver
reached a point where the cost can no longer improve.
Key Features (v1.4.0)
- Slot-Map Problem Structure (faster): Variables and factors are stored in a
slotmap-backed arena and referenced by stable, generationalVarKey/FactorKeyhandles instead of string keys. This gives O(1) access with no hashing or per-key allocation on the hot path, and keeps manifold parameters in contiguousnalgebrastorage thatfaerviews without copying โ minimizing data movement between the two linear-algebra backends. See Performance. - Bundle Adjustment with Camera Intrinsic Optimization: Simultaneous optimization of camera poses, 3D landmarks, and camera intrinsics (10 camera models via apex-camera-models crate) apex-camera-models
- Explicit & Implicit Schur Complement Solvers: Memory-efficient matrix-free PCG for large-scale problems (10,000+ cameras) alongside traditional explicit formulation
- 15 Robust Loss Functions: Comprehensive outlier rejection (Huber, Cauchy, Tukey, Welsch, Barron, and more)
- Manifold-Aware: Full Lie group support (SE2, SE3, SO2, SO3, SE_2(3), SGal(3), Sim(3), Rn) with analytic Jacobians apex-manifolds
- Three Optimization Algorithms: Levenberg-Marquardt, Gauss-Newton, and Dog Leg with unified interface
- Prior Factors & Fixed Variables: Anchor poses with known values and constrain specific parameter indices
- Uncertainty Quantification: Covariance estimation for both Cholesky and QR solvers
- Real-time Visualization: Integrated Rerun support for live debugging of optimization progress
- I/O: Read and write G2O, Toro, BAL format files for seamless integration with SLAM ecosystems apex-io
- High Performance: Sparse linear algebra with persistent symbolic factorization
- Mathematical Cookbooks: Full derivations and explanations for apex-manifolds, apex-camera-models, and apex-io
- Production-Grade: Comprehensive error handling, structured tracing, integration test suite
Quick Start
[]
= "1.4.0"
use Problem;
use BetweenFactor;
use ;
use ;
use dvector;
use HashMap;
Result:
Status: CostToleranceReached
Initial cost: 1.280e+05
Final cost: 2.130e+01
Iterations: 5
Architecture
The workspace root is the apex-solver crate. Sub-crates for manifolds, I/O, and camera models live in crates/:
apex-solver/ # workspace root = apex-solver crate
โโโ src/
โ โโโ core/ # Problem formulation, factors, residuals
โ โโโ factors/ # Factor implementations (projection, between, prior)
โ โโโ optimizer/ # LM, GN, Dog Leg algorithms
โ โโโ linalg/ # Cholesky, QR, Explicit/Implicit Schur
โ โโโ observers/ # Optimization observers and callbacks
โโโ bin/ # Executable binaries
โโโ benches/ # Benchmarks
โโโ examples/ # Example programs
โโโ tests/ # Integration tests
โโโ doc/ # Extended documentation
โโโ crates/
โโโ apex-manifolds/ # Lie groups: SE2, SE3, SO2, SO3, SE_2(3), SGal(3), Sim(3), Rn
โโโ apex-io/ # File I/O: G2O, TORO, BAL formats
โโโ apex-camera-models/ # 8 camera projection models
Core Modules (in src/):
core/: Optimization problem definitions, residual blocks, robust loss functions, and variable managementoptimizer/: Three optimization algorithms (Levenberg-Marquardt with adaptive damping, Gauss-Newton, Dog Leg trust region) with real-time visualization supportlinalg/: Linear algebra backends including sparse Cholesky decomposition, QR factorization, explicit Schur complement, and implicit Schur complement (matrix-free PCG)observers/: Optimization observers and callbacks (Rerun visualization, custom hooks)
Workspace Sub-crates (in crates/):
apex-manifolds: Lie group implementations (SE2, SE3, SO2, SO3, SE_2(3), SGal(3), Sim(3), Rn) with analytic Jacobiansapex-io: File format parsers for G2O, TORO, and BAL formatsapex-camera-models: Camera projection models with analytic Jacobians (10 models)
Low-level Dependencies:
faer/nalgebra: High-performance linear algebra backends
Performance & Data Structure
Apex Solver stores the optimization problem in a slot-map arena. Problem keeps its
variables and residual blocks in slotmap::SlotMaps and returns stable, generational
VarKey / FactorKey handles; per-variable side data (fixed indices, bounds, column
offsets) lives in matching SecondaryMaps.
Why it's faster than the previous string-keyed design:
- O(1) generational access, no hashing. Looking up a variable during residual/Jacobian
assembly is an index + generation check, not a
HashMap<String, _>hash and compare. - No per-key allocation.
VarKey/FactorKeyareCopy8-byte handles; there are noStringkeys to allocate, clone, or compare. - Cache-friendly iteration. Values live in a dense backing array, so assembly sweeps contiguous memory.
- Generational safety. A removed handle can never alias a reused slot โ stale keys
return
Noneinstead of silently pointing at a different variable.
The handles also enable a zero-copy nalgebra โ faer boundary: manifold parameters stay in
contiguous nalgebra column-major storage and are handed to factors as &[f64] slices that
faer views directly (MatRef/MatMut::from_column_major_slice) โ no DVectorโMat
conversion in the inner loop. Combined with a persistent symbolic factorization (built once,
reused every iteration) and lock-free parallel assembly over disjoint buffers (rayon), the
per-iteration hot path is allocation- and copy-free for the manifold data.
โ Full performance benchmarks
Datasets
Datasets are downloaded on demand using the built-in download_datasets tool in the apex-io crate. No Git LFS required.
# List all available datasets and selection numbers
# Download benchmark datasets (all odometry g2o + largest from each BA dataset)
# Download all odometry g2o datasets (2D + 3D)
# Interactive mode (prompts for selection)
Datasets are saved to data/odometry/ (g2o files) and data/bundle_adjustment/ (BAL format).
Available datasets:
- Pose Graph SE2 (2D):
M3500,mit,city10000,ring - Pose Graph SE3 (3D):
sphere2500,parking-garage,torus3D,cubicle - Bundle Adjustment (UW BAL):
ladybug,trafalgar,dubrovnik,venice,final
Workspace Crates
Apex Solver is organized as a Cargo workspace with specialized sub-crates that can be used independently:
| Crate | Description | Docs | Cookbook |
|---|---|---|---|
| apex-manifolds | Lie group manifolds (SE2, SE3, SO2, SO3, SE_2(3), SGal(3), Sim(3), Rn) with analytic Jacobians | README | Cookbook |
| apex-camera-models | 10 camera projection models for bundle adjustment and SLAM | README | Cookbook |
| apex-io | File I/O utilities for G2O, TORO, and BAL formats | README | Cookbook |
Cookbooks
Each sub-crate ships an mdBook cookbook (KaTeX-rendered) that is the mathematical reference for its domain โ derived from the implementation, not restated from papers:
- apex-manifolds โ every group
and operation: exp/log, adjoints, left/right Jacobians and inverses, โ/โ, plus a shared
Conventions page
documenting w-first quaternions and
[ฯ, ฮธ]twist order. - apex-camera-models โ one chapter per model on an eight-section template (Parameters โ Projection โ Inverse Projection โ Point Jacobian โ Intrinsic Jacobian โ Linear Estimation โ Example โ References), with validity conditions merged into the projection sections.
- apex-io โ every public capability by domain: pose-graph formats, ASL/EuRoC, ROS1/ROS2 bags, DDS, CLI tools, and a feature-flag reference.
Build any of them locally:
Using sub-crates independently:
[]
= "0.3.0"
[]
= "0.3.0"
[]
= "0.3.0"
Performance Benchmarks
Detailed benchmark tables comparing apex-solver against Ceres, GTSAM, g2o, factrs, and tiny-solver on 8 pose-graph datasets (SE2/SE3) and 4 BAL bundle-adjustment datasets.
โ Full performance benchmarks
Examples
Usage examples covering pose graph optimization, custom factor implementation, and self-calibration bundle adjustment.
โ Full examples
Technical Implementation
Robust Loss Functions
15 robust loss functions for handling outliers in optimization:
- L2Loss: Standard least squares (no outliers)
- L1Loss: Linear growth (light outliers)
- HuberLoss: Quadratic near zero, linear after threshold (moderate outliers)
- CauchyLoss: Logarithmic growth (heavy outliers)
- FairLoss, GemanMcClureLoss, WelschLoss, TukeyBiweightLoss, AndrewsWaveLoss: Various robustness profiles
- RamsayEaLoss: Asymmetric outliers
- TrimmedMeanLoss: Ignores worst residuals
- LpNormLoss: Generalized Lp norm
- BarronGeneralLoss, AdaptiveBarronLoss: Adaptive robustness
- TDistributionLoss: Statistical outliers
Usage:
use HuberLoss;
let loss = new; // 95% efficiency threshold
problem.add_residual_block;
Optimization Algorithms
Levenberg-Marquardt (Recommended)
- Adaptive damping between gradient descent and Gauss-Newton
- Robust convergence from poor initial estimates
- Supports covariance estimation for uncertainty quantification
- 9 comprehensive termination criteria (gradient norm, cost change, trust region radius, etc.)
Gauss-Newton
- Fast convergence near solution
- Minimal memory requirements
- Best for well-initialized problems
Dog Leg Trust Region
- Combines steepest descent and Gauss-Newton
- Global convergence guarantees
- Adaptive trust region management
Linear Algebra Backends
Four sparse linear solvers for different use cases:
- Sparse Cholesky: Direct factorization of J^T J + ฮปI - fast, moderate memory, best for well-conditioned problems
- Sparse QR: QR factorization of Jacobian - robust for rank-deficient systems, slightly slower
- Explicit Schur Complement: Constructs reduced camera matrix S = B - E Cโปยน Eแต explicitly in memory - most accurate for bundle adjustment, moderate memory usage
- Implicit Schur Complement: Matrix-free PCG solver computing only Sยทx products - memory-efficient for large-scale problems (10,000+ cameras), highly scalable
Configure via LinearSolverType in optimizer config:
config.with_linear_solver_type // For bundle adjustment
config.with_linear_solver_type // For very large BA
Interactive Visualization
Real-time optimization debugging with integrated Rerun visualization using the observer pattern:
use ;
let config = new
.with_max_iterations;
let mut solver = with_config;
// Add Rerun visualization observer (requires `visualization` feature)
let result = solver.optimize?;
Visualized Metrics:
- Time series: Cost, gradient norm, damping (ฮป), step quality (ฯ), step norm
- Matrix visualizations: Hessian heat map, gradient vector
- 3D poses: SE3 camera frusta, SE2 2D points
Run Examples:
# Enable visualization feature and run
Note: The data files (e.g.,
sphere2500.g2o) must be downloaded first. See Datasets โ runcargo run --release -p apex-io --bin download_datasets -- --select 10to get all benchmark datasets.
Zero overhead when disabled (feature-gated).
Learning Resources
Computer Vision Background
- Multiple View Geometry (Hartley & Zisserman) - Mathematical foundations
- Visual SLAM algorithms (Durrant-Whyte & Bailey) - Probabilistic robotics
- g2o documentation - Reference C++ implementation
Lie Group Theory
- A micro Lie theory (Solร et al.) - Practical introduction
- manif library - C++ reference we follow
- State Estimation for Robotics (Barfoot) - SO(3) and SE(3)
Optimization Theory
- Numerical Optimization (Nocedal & Wright) - Standard reference
- Trust Region Methods - Dog Leg theory
- Ceres Solver Tutorial - Practical guide
Acknowledgments
Apex Solver draws inspiration and reference implementations from:
- Ceres Solver - Google's C++ optimization library
- g2o - General framework for graph optimization
- GTSAM - Georgia Tech Smoothing and Mapping library
- tiny-solver - Lightweight nonlinear least squares solver
- factrs - Rust factor graph optimization library
- faer - High-performance linear algebra library for Rust
- manif - C++ Lie theory library (for manifold conventions)
- nalgebra - Geometry and linear algebra primitives
License
Licensed under the Apache License, Version 2.0. See LICENSE for details.