Bundle Adjustment

Bundle adjustment is a technique used in photogrammetry and computer stereo vision to refine a 3D reconstruction built from multiple images of the same scene. Given the observed positions of matched image features, it simultaneously adjusts the 3D coordinates of those features, the cameras' relative motion, and the cameras' optical parameters, minimizing the reprojection error between where the model predicts each feature should appear and where it actually appears in each image. Framed as a nonlinear least squares optimization problem, typically solved with the Levenberg-Marquardt method, bundle adjustment produces a maximum likelihood estimate when the image errors are Gaussian, and it is usually run as the final refining step of feature based 3D reconstruction pipelines. The technique originated in photogrammetry in the 1950s and later became a standard step in computer vision.