Section: New Results

Modeling

Feature Preserving Point Set Surfaces based on Non-Linear Kernel Regression

Participants : Cengiz Oztireli, Gaël Guennebaud, Markus Gross.

In contrast to traditional polygonal meshes, point-based geometries are connectivity free and no topological consistency has to be satisfied through geometry manipulations. This makes point sets one of the simplest and most versatile surface representation since they are particularly well suited to deal with real world acquired data sets, for the streaming of 3D geometries, for the rendering on small devices, etc. Among the wide variety of meshless representations which have been proposed, the ones based on Moving Least Squares (MLS) approximations are ones of the most attractive. However, by definition, MLS based methods play well with smooth data only (Figure 7 ). In collaboration with ETH Zurich, we proposed an efficient and flexible solution combining Local Kernel Regression methods with robust statistics [19] . In particular, our approach allows to achieve real-time performance without any preprocessing, and it allows to take into account all frequency features. This work won the best student paper award at Eurographics 2009.

Figure 7. Illustration of our feature preserving MLS surface definition.

2D Shape Animation

Participants : William Baxter, Barla Pascal, Ken-Ichi Anjyo.

In the context of a collaboration with the Japanese private laboratory OLM Digital R&D, and as part of a broader JST-CNRS project, we developped novel 2D animation tools that allow novice users to easily create animations from pictures, and lets expert users better control more complex animations. Our first contribution, published at the NPAR conference in 2008, proposed extensions and corrections to recent As-rigid-as-possible interpolation techniques  [36] . The main challenge in such techniques resides in establishing correspondences between input drawings though. And it is an essential stage if one wants to make an advanced 2D animation system work in practice. We thus conentrated on this problem, and gave two novel contributions in 2009: we first presented a method for establishing such correspondences between a pair of drawings using Compatible Embeddings  [16] , published in the TVCG journal; we then extended the technique to work with multiple inputs, ending up with a N-way morphing technique [17] , published in the CAVW journal.

Figure 8. Our approach to 2D animation consists in extracting a mapping between a pair of 2D shapes in the form of a compatible embedding, and then morphing between the two shapes with as-rigid-as-possible interpolation.

Dynamic Scene 3D Reconstruction and 3D Collaborative Telepresence

Participants : Benjamin Petit, Jean-Denis Lesage, Yann Savoye, Benoit Bossavit, Jean-Sébastien Franco, Edmond Boyer, Bruno Raffin, Li Guan, Marc Pollefeys.

We are pursuing efforts to build a collaborative, interactive 3D environment based on 3D modeling of subjects from silhouettes observed from multiple, calibrated cameras, in the context of the DALIA project. We are building a collaborative environment where two distant users, each one modeled in real-time in 3D, interact in a shared virtual world. It is based on previous efforts to achieve real-time modeling using the Exact Polyhedral Visual Hulls method [18] , which provides full-body geometric and photometric data on the objects present in the acquisition space. We have built a first version of the interactive system which uses such models as input rendering user's 3D avatars, and for computing interactions an collisions with virtual objects through a physical simulation engine. This plateform was demonstrated at VRST 2008, with transfers over a local area network, and presented in 3DVTCON'09 [28] . In 2009, we have also been working on more advanced 3D methods to enable efficient 3D motion analysis. Such methods are needed for finer collision-based interactions and also allow a more compact transmission of information for telepresence. There are two works in progress toward this goal, yet to be published. The first is based on deformable mesh tracking techniques (PhD work of Yann Savoye). The second uses a probabilistic 3D shape and motion analysis, an ongoing collaboration between IPARLA, the INRIA Perception Team, ETH Zérich and UNC Chapel Hill (USA). This work is submitted to the CVPR 2010 international conference and accepted as publication at the French RFIA 2010 conference.

Figure 9. One of two interactive 3D shape plateforms of our telepresence prototype. Bottom left: the hands of users are reconstructed and used to interact in the same virtual world.

BRDF Modeling

Participants : Romain Pacanowski, Xavier Granier, Christophe Schlick.

The Bidirectional Reflectance Distribution Function (BRDF) describes the appearance of a material by modeling its interaction with light at a surface point. To account for the great variety of physical phenomena involved in light reflection (diffuse reflection, forward or backward scattering, forward or backward specular reflection), standard analytical models usually express the BRDF as a linear combination of several reflection lobes, each dealing with a specific light behavior. We have introduced a new representation [35] that describes an arbitrary isotropic material by a single polymorphic 2D function. This representation has the ability to: (1) mimic standard analytical reflection models, (2) offer high-quality fitting of measured materials with an efficient and guaranteed convergence to a global minimal error, and (3) provide intuitive manipulation of material behaviors (edition or creation from scratch) using a small set of 1D curves.