PhD position
Mesh simplification with structure control
Context
Recent evolutions in Multiview Stereo allow practitioners to measure urban environments at
resolutions that were until now possible only at the scale of individual shapes [1,2]. The related
scientific challenge is now to transform massive and dense meshes into more compact representations
that are structure-aware [3]. Structure is a generic term that refers here to the way the individual shapes
are grouped to form objects, object classes or hierarchies. In an urban context, exploring the structure
can correspond to the modelling of objects at different Levels Of Detail (LOD), see Fig. 1. This
objective can be realized through complex reconstruction algorithms, eg [4], or, more directly, by
surface approximation. Surface approximation is a traditional research topic which consists in
simplifying the input surface of an object or a collection of objects given some quality measures, eg
faithfulness and compactness. The geometry processing community has deeply explored this topic by
considering defect-free meshes as input data [5]. However, this topic has been poorly explored from
real world meshes, and no efficient solution has been proposed so far from meshes generated from
Multiview Stereo.
Figure 1 – Reconstruction of a building from multiview images at different Levels Of Detail
Objectives
Existing methods, eg [6], typically allow the preservation of the object structure at a given scale, but
fails to bring control upon the structure in order to explore the scale spaces of the structure. The aim is
to develop a methodological framework to create generic 3D models from unorganized parametric
shapes, eg planes, which enable to explore the scale-space of an object or an observed scene while
conforming to a given LOD formalism. Contrary to traditional simplification frameworks, the proposed
should rely on a global approach in which geometric and semantic regularities at the city scale must be
taken into account. Among the difficult questions to solve, the candidate will investigate on how to
produce 3D models with geometric and topological guarantees that conform to a given LOD
formalism, and also on how to manage the missing data and heterogeneous sampling. In particular
what type of knowledge can we consider when geometric descriptors are insufficient, and how to
model it?
Keywords
Computer vision, geometry processing, urban scenes, surface approximation, 3D modeling.
Candidate profile
The ideal candidate should have good knowledge in computer vision, 3D geometry and applied
mathematics, be able to program in C/C++, be fluent in English, and be creative and rigorous.
Location
INRIA, Sophia Antipolis Méditerranée, (near Antibes on the Côte d’Azur, France), within the TITANE
research team (https://team.inria.fr/titane/)
Contact
Florent Lafarge ([email protected])
Julien Soula ([email protected])
References
[1] Vu, Labatut, Pons and Keriven. High Accuracy and Visibility-Consistent Dense Multiview Stereo. IEEE
PAMI 34(5), 2012
[2] Lafarge, Keriven, Brédif, and Vu. A Hybrid Multiview Stereo Algorithm for Modeling Urban Scenes, IEEE
PAMI 35(1), 2013
[3] P. Musialski, P. Wonka, D. Aliaga, M. Wimmer, L. Van Gool and W. Purgathofer. ”A survey of urban
reconstruction”. EUROGRAPHICS State of the art reports, 2012.
[4] Verdie, Lafarge and Alliez. LOD Generation for urban scenes. ACM TOG 34(3), 2015
[5] Botsch, Kobbelt, Pauly, Alliez, and Lévy. Polygon Mesh Processing. AK Peters, 2010.
[6] Salinas, Lafarge and Alliez. Structure-Aware Mesh Decimation, CGF 2015.