Skip to Main content Skip to Navigation

Non-rigid correspondences between surfaces embedded in 3D

Abstract : Handling and processing the massive amount of 3D data has become a challenge with countless applications, such as computer-aided design, biomedical computing, interactive games, machine perception, robotics, etc. Geometry Processing is an area of research at the interface between algorithmics, applied mathematics and computer science related to the above applications, that exists since approximately 50 years. It is a large topic of research that includes sub-areas. The problem of shape correspondence (also known as "shape matching") consists in, given a pair of shapes, finding a "good" correspondence between them. For example we may want the correspondence to preserve geodesic distances, or local geometric features.This problem has received a growing interest, in part due to its wide applicability, for example in animation, shape morphing or statistical shape modeling.The functional map framework is a recent tool that has shown many useful properties for shape matching. This approach provides a smooth compact representation of correspondences between shapes, and most constraints over functional maps can be expressed as linear constraints, which allows a least squares formulation of the problem.In this thesis we focus on the problem of shape correspondence, specifically using functional maps. In Chapter 1 we introduce basic notions and notations that will be used throughout the thesis, related to continuous and discrete surfaces, the Laplace-Beltrami operator, the problem of non-rigid shape matching, and the standard functional map computation pipeline.In Chapter 2 we notice that functional maps that are induced by point-to-point maps should satisfy point-wise product preservation constraints. We apply this observation to shape descriptors in order to improve the previous classical constraints on functional maps. This leads to an approach that allows to extract more information from existing constraints and results in better correspondences, particularly when the number of independent descriptors is small.In Chapter 3 we build on the previous remark, but this time in the situation where we already have a functional map that was computed by an existing method. We notice that the point-wise product preservation can also be used to extend the domain over which the given functional map can transfer functions. We show that this allows to improve the accuracy of function transfer.In Chapter 4 we extend the approach proposed in Chapter 3 by noticing that instead of using point-wise function products, the point-wise composition by any fixed operator should also be preserved. We use a neural network that optimizes the approximation of a given function that we want to transfer, as a point-wise function of some basis functions that we already know how to transfer using a given functional map. We then describe how to apply this trained network to the image of the basis functions to construct the image of the function that we want to transfer. We show preliminary results that suggest that this method can lead to significant improvement for function transfer.Finally, in Chapter 5 we mention other topics studied during the thesis, that are unrelated to non-rigid shape matching.
Complete list of metadata

Cited literature [107 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Thursday, April 4, 2019 - 8:03:00 PM
Last modification on : Monday, February 3, 2020 - 1:09:10 AM


Version validated by the jury (STAR)


  • HAL Id : tel-02074764, version 2



Dorian Nogneng. Non-rigid correspondences between surfaces embedded in 3D. Image Processing [eess.IV]. Université Paris-Saclay, 2018. English. ⟨NNT : 2018SACLX109⟩. ⟨tel-02074764v2⟩



Record views


Files downloads