HAL will be down for maintenance from Friday, June 10 at 4pm through Monday, June 13 at 9am. More information
Skip to Main content Skip to Navigation

Processing Mesh Surface Signals for Appearance Modeling

Abstract : Recent technological developments, in particular mixed reality media, combined with a pressing need for telepresence applications, give rise to a growing demand for 3D content that is captured directly from real data.Such content can then be edited and rendered in an interactive manner from any given viewpoint, producing an immersive experience for the user.The modeling process usually involves the separate reconstruction of the shape and the appearance, (or texture), of a scene, given a set of 2D photographs. A faithful representation of the appearance is paramount for a realistic rendering.Yet, it still faces many challenges, that we wish to address in this thesis.There is a wide range of methods used for computing, storing and rendering appearance information.Considering the need for efficient storage and streaming capabilities of such models, we choose to focus on the most compact, and probably the most widely used, type of representation: a 3D mesh with a color texture.We contribute to various aspects of appearance modeling in this context, including sampling,compression, and super-resolution of appearance, and surface denoising.First, the amount of available appearance information is not uniform on the surface.Yet, standard parameterizations of the surface, i.e. texture maps, can only sample it uniformly.Using the inner parameterization of the mesh, we propose an adaptive sampling strategy that isable to more efficiently capture available information.Then, because the appearance model we use cannot be compressed efficiently, we equip it witha lossy compression method, that makes it more efficient than image textures in terms of bit-rate vs. visual quality.Geometric noise is one of several sources of noise that make the problem of appearance modeling particularly challenging.Considering the success of convolutional networks for image denoising,we contribute to the field of mesh denoising, with the first end-to-end deep-learning approach for normals denoising, with a graph convolution network.Finally, we tackle the problem of appearance super-resolution, with the goal of producing a detailed appearance from a set of low-resolution images.We present a novel approach, with a data-driven method that is trained to enhance appearance in image space, rather than in a shared texture space.
Complete list of metadata

Contributor : Abes Star :  Contact
Submitted on : Monday, March 21, 2022 - 3:16:18 PM
Last modification on : Friday, March 25, 2022 - 9:40:32 AM


Version validated by the jury (STAR)


  • HAL Id : tel-03615455, version 1



Matthieu Armando. Processing Mesh Surface Signals for Appearance Modeling. Computer Vision and Pattern Recognition [cs.CV]. Université Grenoble Alpes [2020-..], 2021. English. ⟨NNT : 2021GRALM044⟩. ⟨tel-03615455⟩



Record views


Files downloads