A Stable Viewing Strategy for Rotation Normalization Free 3D Model Retrieval

This paper proposes a viewing strategy for multi-view 3D model retrieval based on scale-invariant feature transform (SIFT) descriptors, which could make the retrieval process free from the rotation normalization of the model. To sustain a stable descriptor set invariant to the rotation of 3D models, the affine invariant characteristic of SIFT is analyzed to find a better view point set on the viewing sphere. The geometric features of Platonic Solids are discussed to select a better viewing model for 3D model retrieval based on SIFT descriptors. To prove the geometric advantages of icosahedron, experiments are conducted on the publicly available 3D model criterion Princeton Shape Benchmark (PSB) to compare the stability of different SIFT features sets, which were generated by the cube, dodecahedron and icosahedron viewing models respectively. The experimental results show that the feature set of 3D models based on icosahedron acquires not only higher stability but also smaller number of viewing points compared with other Platonic solids.