Keypoints and Local Descriptors of Scalar Functions on 2D Manifolds

Abstract

International audienceThis paper addresses the problem of describing surfaces using local features and descriptors. While methods for the detection of interest points in images and their description based on local image features are very well understood, their extension to discrete manifolds has not been well investigated. We provide a methodological framework for analyzing real-valued functions defined over a 2D manifold, embedded in the 3D Euclidean space, e.g., photometric information, local curvature, etc. Our work is motivated by recent advancements in multiple-camera reconstruction and image-based rendering of 3D objects: there is a growing need for describing object surfaces, matching two surfaces, or tracking them over time. Considering polygonal meshes, we propose a new methodological framework for the scale-space representations of scalar functions defined over such meshes. We propose a local feature detector (MeshDOG) and region descriptor (MeshHOG). Unlike the standard image features, the proposed surface features capture both the local geometry of the underlying manifold and the scale-space differential properties of the real-valued function itself. We provide a thorough experimental evaluation. The repeatability of the feature detector and the robustness of feature descriptor are tested, by applying a large number of deformations to the manifold or to the scalar function

Toward user-specific tracking by detection of human shapes in multi-cameras

“…We consider animated meshes as testing data, where groundtruth vertex indices are available. The error metric is the geodesic distance between predicted vertices and groundtruths, and we compare with MeshHOG [30] which is the extension of image-based histogram of oriented gradient (HOG) on surface manifolds. Fig.…”

“…The calculation of f is suggested to be dimensionally independent. We therefore avoid descriptors that requires normalization, like MeshHOG [30], or SHOT [28], and resort to comparison features used in [10,22].…”

Algorithm for Registering Point Cloud Data with Quadratic Surfaces

“…Zaharescu et. al., [4] and J. Stuckler et. al., [5] achieved detection of feature points in the non-uniform mesh surface by performing multi-scale operations on the curvature and the normal vector, respectively.…”

Retrieval and classification methods for textured 3D models: a comparative study

“…The attention towards texture properties has grown considerably over the last few years, as demonstrated by the number of techniques for the analysis of geometric shape and texture attributes that have been recently proposed [33,46,54,64,75,80]. Since 2013, a retrieval contest [9] has been launched under the umbrella of the SHREC initiative [76] to evaluate the performances of the existing methods for 3D shape retrieval when dealing with textured models.…”