We present an original algorithm to construct Voronoi tessellations on surfaces from a set of depth maps. Based on a local graphbased structure, where each local graph spans one depth map, our algorithm is able to compute partial Voronoi diagrams (one per scan), and then to merge/update them into a single and globally consistent Voronoi diagram. Our first results show that this algorithm is particularly promising for improving the sampling quality of massive point clouds or for reconstructing very large-scale scenes, with low and manageable memory consumption.