Fusing Multiple Features for Shape-based 3D Model Retrieval

Fusing multiple features is a promising approach for accurate shape-based 3D Model Retrieval (3DMR). Most of the previous algorithms either simply concatenate feature vectors or sum similarities derived from features. However, ranking results due to these methods may not be optimal as they don’t exploit distributions, i.e., manifold structures, of multiple features. This paper proposes a novel 3DMR algorithm that effectively and efficiently fuses multiple features. The proposed algorithm employs a Multi-Feature Anchor Manifold (MFAM) that approximates multiple manifolds of heterogeneous features with small number of “anchor” features. Given a query, ranks of 3D models are computed efficiently by diffusing relevance on the MFAM. Distance metrics of heterogeneous features are fused during the diffusion for better ranking. Experiments show that our proposed algorithm is more accurate and much faster than 3DMR algorithms we have compared against.