Correspondence and Matching
Contributors
- Vasyl Mykhalchuk, Frederic Cordier, Hyewon Seo (University of Strasbourg)
Abstract
A lot of attention has been paid to feature extraction and feature matching problems during the past decade. Here, we aim at solving a different and relatively unexplored problem: automatic transfer of user-defined feature points, independently from geometric saliency. This technique is useful when a user is interested in characterization and selection of points that are not necessarily geometrically significant (such as some anatomical landmarks defined on non-prominent regions), on meshes of similar shape (meshes coming from the same class of objects, for instance).
Briefly, our approach is based on surface graphs and their matching (as shown in the image above). We build a graph G_S on the source mesh, whose nodes are a set of automatically selected geometric feature points and whose edges are weighted by the geodesic distances. For the extraction of geometric features, we have defined a local shape descriptor, which is invariant to isometry and insensitive to the mesh discretization as much as possible. Then, given a user-specified landmark on the source mesh, we first build what we call minimal graph G_M, a subgraph of G_S. The graph G_M has two main properties: (1) it uniquely defines the user-provided landmark, (2) it is as small as possible in terms of number of nodes and geodesic distances. Next, given a target mesh, we select a set of points with the local shape signatures similar to the points from graph G_S. From these feature points we compute the graph G_T by connecting the points which are within the maximum geodesic radius of G_S. Having both graphs, G_M and G_T, we use existing technique to find a subgraph of G_T, that best matches with G_M.
Based on this feature detection work, we have developed a new technique that makes use of deformation properties between animated meshes for finding their spatial correspondences. Given a pair of animated meshes M and M' exhibiting semantically similar motions, we detect a sparse set of feature points on each mesh and compute spatial correspondences between them, so that points with similar motion behavior are put in correspondence. At the core of our technique is a new, dynamic feature descriptor named AnimHOG, which encodes local deformation characteristics. AnimHOG is obtained by computing the gradient of a scalar field inside the spatiotemporal neighborhood of a point of interest, where the scalar values are obtained from the deformation characteristic associated with each vertex and at each frame.
