SLAM is the generic name given to the class of methods allowing to incrementally build a 3D representation of an environment while simultaneously using this map to localize a mobile system evolving within this environment. Though quite a mature field, several scientific problems remain open and particularly the reduction of drift. Drift is inherent to SLAM since the task is fundamentally incremental and er- rors in model estimation are cumulative. In this paper we suggest to take advantage from sparse but accurate knowl- edge of the environment to periodically reinitialize the sys- tem, thus stopping the drift. As it may be of interest in a Augmented reality context, we show this knowledge can be propagated to past estimations through bundle adjust- ment and present three different strategies to perform this propagation. Experiments carried out in an urban environ- ment are described and demonstrate the efficiency of our approach.