In this letter, we propose an approach for merging three-dimensional maps represented as pose graphs of point clouds. Our method can effectively deal with typical distortions affecting simultaneous localization and mapping-generated maps. Traditional map merging techniques that use a single rigid body transformation to relate the reference frames of different maps. Instead, our approach achieves more accurate results by eliminating the inconsistencies resulting from distortions affecting the inputs, and can succeed in those situations where traditional approaches fail for substantial deformations. The core idea behind our solution is to localize the robot in a reference map by using the data from another map as observations. We validated our approach on publicly available datasets, and provide quantitative results that confirm its effectiveness on challenging instances of the merging problem.