Obtaining 3D models from long image sequences is a major issue in computer vision. One of the main tools used to obtain accurate structure and motion estimates is bundle adjustment. Bundle adjustment is usually performed using non-linear Newton-type optimizers such as Levenberg-Marquardt which might be quite slow when handling a large number of points or views.

We investigate an algorithm for bundle adjustment based on quasi-linear optimization. The method is straightforward to implement and relies on solving weighted linear systems obtained as simple functions of the input data. Important features are that (i) the original cost function is preserved, (ii) different projection models, either calibrated or not, are handled in a unified framework and (iii) any number of views and points as well as missing data can be handled.

Experimental results on simulated and real data show that the algorithm is as accurate as standard techniques while requiring less computational time to converge.