Geometry-aware point cloud clustering for spherical-component aggregate modeling

Abstract

Abstract This paper proposes a method for obtaining independent mesh models of individual components from a point cloud representing an aggregate. An aggregate consists of a collection of small, similar components, such as individual grapes in a bunch. Typical shape reconstruction creates a rough shape of the entire bunch, but fails to recover individual components from the bunch due to occlusion and missing points for shapes. To achieve this type of modeling, we assume that each component can be approximated as a spherical shape. Leveraging this assumption, we develop geometry-aware clustering that identifies and segments individual components from the aggregate. During this procedure, we search for the optimal position and size of a predefined aggregate component that best fits the cluster. When overlapping components are detected, the corresponding clusters are merged. We demonstrate the effectiveness of the proposed method by applying it to several types of aggregates, such as grapes and tomatoes.