3D Scene Segmentation with a Shape Repository

Abstract

3D scene segmentation is important but difficult especially when one object makes contact with another in scenes. This paper presents a new algorithm for automatic object extraction with the help of a shape repository, which is an effective approach applicable to 3D meshes without scene graphs. In the new algorithm, connected components are first computed and taken as initial clusters. Then, based on the global shape similarities between the supposed mergence of two shapes and any shape from the shape repository, we make a decision as to whether two close shapes should be merged or not. Our iterative merging scheme is performed until none of the close shapes can be merged. Experiments show that this method is very efficient on some indoor scenes.