MpLBP: A point-based representation for surface pattern description

Abstract

The Local Binary Pattern (LBP) is a very popular pattern descriptor for images that is widely used to classify repeated pixel arrangements in a query image. Several extensions of the LBP to surfaces exist, for both geometric and colorimetric patterns. These methods mainly differ on the way they code the neighborhood of a point, balancing the quality of the neighborhood approximation with the computational complexity. For instance, using mesh topological neighborhoods as a surrogate for the LBP pixel neighborhood simplifies the computation, but this approach is sensitive to irregular vertex distributions and/or might require an accurate surface re-sampling. On the contrary, building an adaptive neighborhood representation based on geodesic disks is accurate and insensitive to surface bendings but it considerably increases the computational complexity. Our idea is to adopt the kd-tree structure to directly store a surface described by a set of points and to build the LBP directly on the point cloud, without considering any support mesh. Following the LBP paradigm, we define a local descriptor at each point that is further used to define a global statistical Mean Point LBP (mpLBP) descriptor. When used to compare shapes, this descriptor reaches state of the art performances, while keeping a low computational cost. Experiments on benchmarks and datasets from real world objects are provided altogether with the analysis of the algorithm parameters, property and descriptor robustness.