Shape segmentation by hierarchical splat clustering

This paper presents a novel hierarchical shape segmentation method based on splats for 3D shapes. The major contribution is to propose a new similarity metric based on splats, which combines patch-aware similarity and part-aware similarity adaptively. An optimized L 2 , 1 metric for VSA (variational shape approximation) method is used to get splats first, and such adaptive similarity metric is used to generate a hierarchy of components automatically through adaptive cluster. As a result, a binary tree is used to represent the hierarchy, in which low level segments are patch-aware regions while high level segments are part-aware components. Therefore, the combination and decomposition relations are clear between segments. Our method is designed to handle arbitrary models, such as CAD model, scanned object, organic shape, large-scale mesh and noisy model. A large number of experiments demonstrate the efficiency of our algorithm. Graphical abstractDisplay Omitted HighlightsAn optimized L 2 , 1 metric is used for VSA method.A novel patch-aware similarity metric is proposed.We improve SDF calculation by using anisotropic smoothing.The patch and part aware similarities are adaptively combined into a uniform metric.A hierarchy of segmentations are obtained with our hierarchical splat clustering.