Shape segmentation is a very important topic in shape analysis. In this paper, a novel hierarchical shape segmentation method based on splats for 3D shapes is proposed for both patch-aware and part-aware purposes.

The authors present a hierarchical splat clustering framework. Using the splats as the initial clusters, the authors calculate an improved variational shape approximation (VSA) with an L^(2,1) metric. The patch-aware similarity metric and part-aware similarity metric are then defined and combined for merging the neighboring clusters hierarchically to produce a hierarchy of regions. The method generates the hierarchical clustering output as a binary tree denoting the different levels of segmentation. A boundary smoothing would be applied in the post-processing step to improve the segmentation quality in the implementation. Extensive experiments on different models, such as CAD model, scanned object, organic shape, large-scale mesh, and noisy model, show impressive segmentation results. The runtime table in the paper demonstrates its great efficiency.

The main contribution of this paper is the presentation of a new hierarchical clustering segmentation algorithm for shape segmentation. The beauty of this method is that it combines patch-aware similarity and part-aware similarity. As the authors state in the paper, “One limitation of our approach is that it [is] not suit[ed] for all semantic parts adher[ing] to our part characterization.”