Non-rigid 3D Model Retrieval Using Set of Local Statistical Features

Abstract

Various algorithms for shape-based retrieval of non-rigid 3D models, with invariance to articulation and/or global deformation, have been developed. A majority of these algorithms assumes that 3D models have mathematically well-defined representations, e.g., closed, manifold mesh. These algorithms are thus not applicable to other types of shape models, for example, those defined as polygon soup. This paper proposes a 3D model retrieval algorithm that accepts diverse 3D shape representations and is is able to compare non-rigid 3D models. The algorithm employs a set of hundreds to thousands of 3D, statistical, local features to describe a 3D model. These features are integrated into a feature vector per 3D model by using bag-of-features approach for efficiency in comparing 3D models and for invariance against articulation and global deformation. Experimental evaluation showed that the algorithm performed well for non-rigid 3D model retrieval.