Densely sampled local visual features on 3D mesh for retrieval

The Local Depth-SIFT (LD-SIFT) algorithm by Darom, et al. [2] captures 3D geometrical features locally at interest points detected on a densely-sampled, manifold mesh representation of the 3D shape. The LD-SIFT has shown good retrieval accuracy for 3D models defined as densely sampled manifold mesh. However, it has two shortcomings. The LD-SIFT requires the input mesh to be densely and evenly sampled. Furthermore, the LD-SIFT can't handle 3D models defined as a set of multiple connected components or a polygon soup. This paper proposes two extensions to the LD-SIFT to alleviate these weaknesses. First extension shuns interest point detection, and employs dense sampling on the mesh. Second extension employs remeshing by dense sample points followed by interest point detection a la LD-SIFT Experiments using three different benchmark databases showed that the proposed algorithms significantly outperform the LD-SIFT in terms of retrieval accuracy.