Modeling and solving bundle adjustment problems
Résumé
We present a modeling of bundle adjustment problems in Julia, as well as a solver for nonlinear least square problems (including bundle adjustment problems). The modeling uses NLPModels
Julia’s library and computes sparse Jacobians analytically. The solver is based on the LevenbergMarquardt algorithm and uses QR or LDL factorization, with AMD or Metis permutation algorithm.
The user can choose to use normalization and line search. Our experimental results contain comparison
of the several versions of the solver and comparison with Scipy’s least square function and Ceres solver
on the test problems given in [26]. We show that our solver is quite competitive with Scipy’s solver
and Ceres solver in terms of convergence, and that it is in average two times faster than Scipy’s solver
and three times slower than Ceres. However, the advantage of our solver is that it is coded is Julia
and thus allows the user to run it in several precisions in a very efficient way, in order to gain time
and energy (in small precisions) or accuracy (in big precisions).
Fichier principal
2020 Cahier du GERAD ANGLA bundle adjustment.pdf (879.27 Ko)
Télécharger le fichier
Origine | Accord explicite pour ce dépôt |
---|