bundle adjustment
**Bundle adjustment** is the **joint nonlinear optimization that refines camera poses and 3D landmark positions by minimizing total reprojection error across all observations** - it is the gold-standard backend step for high-accuracy 3D reconstruction and SLAM consistency.
**What Is Bundle Adjustment?**
- **Definition**: Global least-squares optimization over pose and structure variables.
- **Objective**: Minimize distance between observed feature points and projected 3D landmarks.
- **Variables**: Camera intrinsics or extrinsics plus 3D point coordinates.
- **Optimization Style**: Iterative methods such as Levenberg-Marquardt on sparse Jacobians.
**Why Bundle Adjustment Matters**
- **Global Accuracy**: Corrects drift and local linearization errors accumulated in front-end tracking.
- **Map Consistency**: Produces coherent geometry and trajectory in one solution.
- **High-Precision Applications**: Essential for metrology-grade reconstruction and mapping.
- **Benchmark Standard**: Reference backend for evaluating pose and structure quality.
- **Loop Closure Integration**: Effectively distributes global constraints after revisits.
**BA Components**
**Observation Graph**:
- Tracks which camera observes which landmark.
- Defines sparse optimization structure.
**Residual Model**:
- Reprojection residuals per feature correspondence.
- Optional robust losses handle outliers.
**Sparse Solver**:
- Exploits block-sparse Jacobian for scalability.
- Balances speed and numerical stability.
**How It Works**
**Step 1**:
- Initialize poses and landmarks from front-end matches and triangulation.
**Step 2**:
- Iteratively optimize all variables to minimize reprojection error until convergence.
Bundle adjustment is **the precision-tightening backend that makes maps and trajectories globally coherent and metrically reliable** - despite its compute cost, it remains indispensable for high-quality SLAM and SfM systems.