Bundle Adjustment

Jointly fine-tunes all camera poses and 3D point coordinates to minimize reprojection error.

Bundle adjustment traces back to photogrammetry in the 1950s; “bundle” refers to the bundle of light rays traveling from each 3D point to a camera’s optical center. Given feature points matched across multiple images, it treats every camera’s pose — sometimes intrinsics and distortion too — and every 3D point’s coordinates as unknowns, and minimizes reprojection error: the sum of squared distances between where each 3D point projects under the current parameters and where it was actually observed in the image. When image noise is zero-mean Gaussian, this is equivalent to maximum likelihood estimation. It is typically solved with nonlinear least-squares methods like Levenberg–Marquardt, exploiting the problem’s sparse structure for speed. Bundle adjustment is the core step of structure from motion (SfM) and the back end of visual SLAM — COLMAP reconstruction and ORB-SLAM’s local and global optimization are both running BA — with common solver libraries including Ceres, g2o, and GTSAM.

ExampleWalking around a table taking dozens of photos with a phone: SfM first roughly estimates each photo’s camera pose and triangulates a sparse point cloud, then runs one pass of BA to jointly refine everything, dropping the reprojection error and making both the point cloud and camera trajectory more accurate.

Also called

BA

Related

Structure from Motion · Reprojection Error · Simultaneous Localization and Mapping · SLAM Front-end / Back-end · Factor Graph Optimization · Ceres Solver

Sources

Bundle adjustment - Wikipedia