We propose a flexible method to extract a set of segments from a 3D point cloud that are relevant across several scales and optimal in a planar sense. Since planar geometric primitives are ubiquitous, especially in man-made scene, their accurate detection is crucial for an abstract representation of point-based 3D data. In this paper, we introduce a new hierarchical graph representation in which each node represents a region at a given scale. The proposed graph is initialized with multiple segmentations performed at different scales and then reduced by collapsing groups of nodes. Each resulting group of nodes defines a meaningful segment and is obtained through an optimization that balances number of extracted segments and accuracy with respect to the input data in a planar sense. The output graph is a compact abstraction of the input cloud into multiple, possibly overlapping, segments, each relevant at a certain scale. The edges of the graph connect nodes whose segments overlap across different scales, thus allowing to represent both detailed and approximating parts of the scene.