Despite 3D object recognition being an ongoing research field for many years, state-of-the-art methods still face problems in real-world situations with clutter, occlusion or non-textured objects. To overcome these problems, recent approaches use multi-view setups exploiting beneficial vantage points of the environment. Minimizing the assumptions on the scene and objects of interest made by these systems, we present an efficient online multi-view method, which integrates information of the captured environment merging individual single-view recognition outputs. Our method achieves state-of-the-art results for the Willow dataset at reduced computational time. Further evaluations on the more challenging TUW dataset show an increase in f-score and object pose accuracy over the number of observations.