Modeling and Acquisitions

Permanent participants

• 2 Professors: Dominique Bechmann et Jean-Michel Dischler
• 1 Associate researcher : Hyewon Seo
• 4 Assistant professors : Rémi Allègre, David Cazier, Pierre Kraemer et Basile Sauvage
• 2 Research engineers: Olivier Génevaux et Sylvain Thery

Others participants

• 1 Collaborateur extérieur: Frédéric CORDIER (MC UHA Mulhouse, 09/2008-)
• 1 engineer : Cyril KERN (ANR VORTISS puis STREP PASSPORT 09/2007-05/2010)
• 4 Post-doctoral: Olivier GUILLOT (ATER, 09/2008-08/2009), Younis HIJAZI (STREP PASSPORT, 02/2009-12/2009), Thomas JUND (ATER, 10/2010-08/2011), Frédéric LARUE (Région Alsace, 12/2008-08/2009)
• 2 PhD students: Lionel UNTEREINER (Ministère, 10/2010-), Kenneth VANHOEY (Ministère, 10/2010-)
• 6 Former PhD students: Dobrina BOLTCHEVA (Région Alsace-IRCAD, 10/2003-10/2007), Marc FOURNIER (Bourse du Canada, 10/2004-11/2008), Thomas JUND (Ministère, 10/2007-09/2010), Pierre KRAEMER (Ministère, 10/2005-11/2008), Frédéric LARUE (RIAM AMI3d, 10/2005-11/2008), Benjamin SCHWARZ (Contrat, 10/2004-08/2009).

Digital human modeling

Manufacturers that design products today base their size related decisions mostly on 1-dimensional anthropometric data. This method is not precise enough and not suited for the mass production, nor for customized design. If digital human modeling techniques are to be meaningful in this context, they must be able to produce correctly sized 3-dimensional body models and be efficient enough to be used in an interactive runtime setting with minimum user intervention. Computer aided modeling of human body that satisfies such requirements is a difficult task, even with the most sophisticated software available.

One of the most challenging research areas in this context is in developing a robust methodology on automatically building realistic 3D human models that satisfy given size or attribute-related constraints in real-time performance.

Parametrically controllable shape and motion

With the above mentioned target applications considered, the main goal of this research is to develop methods to create desired models on the fly through a set of user-controllable parameters. At this time, a variety of 3D reconstruction methodologies are available that can capture shapes that exist in the real world. Our initial models or examples rely on the 3D shape capture technology. The initial models from capture technology are then organized to serve as input into the shape interpolator. Each example body scans are converted into a morphable and animatable model through the conformation of a template model. By builing statistically learning models between the control parameters and the target geometry, we obtain highly controllable yet realistic digital human models.

We focus on : (1) The development of a method for accurately learning the variation from dataset, especially to handle multiple modalities such as morphologidal identity, and poses or dynamic variability. Given sufficiently large datasets, the model will capture both static (identity-dependent) and dynamic (movement-dependent) shape variation in a correlated fashion, enabling a continuous range of evaluation. (2) The extension of the parametric deformation model to the motion data. The goal is to generate desired motion through multi-way blending of realistic motion from motion capture data.

Registration using Dynamic Data

Isocontours of degree of deformation of a sphere

Registration, or optimal alignment of two shapes in arbitrary configurations, is a fundamental problem in shape acquisition and modeling. Given two shapes, often called model and data, the goal is to determine a transformation that optimally aligns the model with data. This process is necessary in order to be able to compare or integrate the data obtained from different measurements. Measurement data from human, the most frequently observed subject of imaging devices, set several technical challenges: (1) Reliable correspondence computation becomes a complex problem to solve. It is quite obvious that the difference between geometric features as well as shapes could be far too large due to the inherently large variation among different individuals. (2) Human bodies are not only just flexible but also highly mobile, to drastically change their spatial arrangement or pose.

Taking a step beyond the existing methods that use static shape information for feature computation, we aim for novel approaches by exploiting large redundancy of information that resides in dynamic or movement data. We have acquired subjects’ movement data and developed deformation analysis method so as to characterize dynamic features on them [KCWK10][COHK09].

Degree and direction of principal deformation of bending knees (30 and 60 degrees of flexion)

We are currently devising a new registration technique that makes use of this rich set of information to guarantee reliable correspondence. This work is being pursued in the framework of the French national project SHARED (Shape Analysis and Registration of People using Dynamic Data).

Perspectives

Representation of 3D scanned objects for real-time visualization

3D model reconstruction from multiple data sources

We have designed an efficient 3D acquisition pipeline for constructing 3D geometric models from scanned objects, that reproduce both the shape and appearance of the objects at different levels of detail. Unfortunately, these techniques do not easily scale to hundreds of millions or billions of sample points. This is first due to the complexity of the algorithms and/or their memory consumption. Out-of-core and streaming techniques that operate on data-structures stored in mass storage memory are only suitable for localized treatments or processing at low scale, at the expense of a reduced robustness to defect-laden data incorporating noise or missing parts.

Another factor that hampers scalability is the frequency of user intervention throughout the 3D acquisition pipeline. Despite the significant advances regarding the automation of the algorithms, user interaction is still often required to fit parameters or to perform corrections. Sometimes it is required to restart the whole process from scratch, which is very penalizing when performing complex operations on large data sets. A major difficulty in reconstructing geometric models is the lack of prior information about the properties of the scanned objects: structure, global and local shape, components, etc. Only 3D-scanned data are generally taken into account by existing reconstruction algorithms at this step of the pipeline.

To overcome these issues, we aim at developing new techniques taking advantage of the complementarity of multiple data sources for the reconstruction of geometric models: scans, photographs, 3D models with similar shape and/or appearance, sketches, etc. Our goal is to devise integrated methods built on the analysis of these different kinds of data, making it possible to capture the main features of the scanned objects, that will then be used to localize and guide the reconstruction process with minimum user interaction.

3D model representation for real-time visualization

The huge size and complexity of models represented using standard triangle mesh data-structures involve numerous storage and visualization issues. Our research includes development of alternative representations offering both compactness and suitability for real-time realistic rendering, as well as suitability for some edition operations (virtual sculpture, simulation of natural phenomena).

Representations based on the displacement mapping technique and hierarchical voxel grids efficiently take advantage of the computational power of current graphic processors. These data-structures both require large storage space in graphic memory for complex 3D models, that is not necessarily justified by the local geometry. In order to visualize complex models with limited memory resources, we are developing a new approach of procedural modeling for surfaces based on displacement maps and procedural generation techniques.