Hello, I’m Thibaut !
PhD in Computer Graphics specialized in the research area of virtual characters animation. I defended my PhD on 9th of December 2013 under the supervision of Nicolas Courty and Sylvie Gibet (Expression team in Brittany [France]).
I am currently a lecturer at the University of Fribourg (Switzerland). I work on the analysis and synthesis of human movement in the fields of Computer Science and Cognition (movement analysis, motor learning, relationship to others and analysis of perceptual strategies using eye tracking). My main work focuses on the conception, implementation and assessment of HMI models for the learning of motor and perceptual-motor skills. To achieve that, several sub-issues are addressed such as kinematic motion analysis (body and gaze), the conception of efficient IAs dedicated to Human-Avatar interactions, the design of experimental protocols to assess these models as well as the analysis of data resulting from the experiments conducted.
What are the common spatio-temporal features between the different gestures performed by expert sport players. Traditionally, research on time series focus on (1) feature selection or feature extraction for classification, (2) on similarities between two time series either on the whole movement (structure-based representation) or on part of movement (shape-based representation), (3) computation of time series linear or non linear correlations. The main issues addressed here focus on (1) time series matching, (2) local similarity selections.
Action Anticipation: where does the expert focus ? On what the AI relies on ? Many studies deal with where experts focus their gazes to anticipate the aims of opponents. It is for example the case in soccer with the penaly kick or in tennis with the service. We analyse here the differences between the character limbs focused by experts and the features extracted by AI.
Judgement of movements. One way to analyse the quality of a movement is to compare it with the movement of an expert. One of the key issue concerns the non adequation between the computer science results and the human ones. Experts rely on subjective information as biomechanical information or "harmony of the movement". The addressed issues here concern (1) the quantification of the subjective part in the human judgement and (2) the implementation of AIs able to reproduce the human judgements.
The spatio-temporal relationship during the progress of the slapshot action between the shooter and the goalkeeper is very close. A few ms of delay on the part of one or the other can change the outcome of the action. One of the advantage given by VR is to control one virtual opponent to interact with the sport player. We propose here a real time VR simulator able to analyse and give a learning to improve sensory-motor skills of a hockey player.
Penalty kick simulator designed to 1. assess the penalty scoring abilities of young and adult confirmed players, and 2. improve their performance using personalized training based on an optimization algorithm. The movements and the dive of the virtual goalkeeper are based on the movements of the kicker during the run-up to the ball. The difficulty of the task is constantly adapting to the performance and progress of each individual player. The movements / animation of the virtual goalkeeper are based on the real, motion-captured movements of a professional goalkeeper.
Virtual reality allows to offer very specific stimuli such as augmented feedback in real time and in interaction with the participant. The topic deals with a new learning model based on HMMs generating a virtual display of the superimposition of oneself (self-modeling) with a virtual expert (expert-modeling). The contribution concerns the participant's self-control of the time dimension in the visualization of the feedback. It is the learner who chooses the speed and the progression of the feedback via his movement. Our model is rigorously validated by a statistical validation and an experiment concerning the mawashi-geri in karate.
The main goal of this study was to assess the use of superimposition of self and expert avatars for motor control skill learning with partial movement. Regarding the contrasted methods and results obtained by previous studies in VR, this study allows to confirm the advantages of using this type of feedback. Most of previous studies in VR assessed this feedback on complex movement evolving the whole body, sometimes with large moves (as for example with dance). Moreover, few studies allow highlighting the advantage of this feedback in comparison with research in self and expert modeling.
The objectives of this study are (1) to assess the use of feedback in 3D (in comparison with video feedback) for motor learning specifically on expert movement reproduction, and (2) to compare the different methods of evaluation used by human science movement (judges evaluation, kinematic features), with the ones proposed by computer animation research areas (dynamic time warping).
Knowing what people look at and understanding how they analyze the dynamic gestures of their peers is an exciting challenge. In this context, we propose a new approach to quantifying and visualizing the oculomotor behavior of viewers watching the movements of animated characters in dynamic sequences. Using this approach, we were able to illustrate, on a 'heat mesh',the gaze distribution of one or several viewers, i.e.,the time spent on each part of the body, and to visualize viewers' timelines,which are linked to the heat mesh. Our approach notablyprovidesan 'intuitive' overview combining the spatial and temporal characteristics of the gaze pattern, thereby constituting an efficient tool forquickly comparing the oculomotor behaviors of different viewers. The functionalities of our system are illustrated through two use case experiments with 2D and 3D animated media sources, respectively validated by a statistical validation and an experiment concerning the penalty in soccer.
The challenge proposed by this study is to reconsider the entire animation pipeline for data-driven character animation. By observing that significant loss of information and precision occur in the traditional animation pipeline (skeleton reconstruction from markers, rigging and retargeting), our goal is to directly control at interactive framerates articulated meshes from a low number of positional constraints. Our method builds on top of efficient deformation techniques and proceeds as follows: an original mesh is embedded into a coarse volumetric control lattice which contains simplified information from the initial reference mesh, skeleton elements and marker locations. An iterative method is applied on this structure which preserves the geometry details, the bones lengths, and the associated joint limits. We show the ability of our approach to animate and interactively deform high resolution models from a low-number of markers while retaining the subtleties of the motion. It notably allows to entirely skip the tedious rigging phase
Use of the discrete Laplacian operator to reconstruct the missing marker position data. Each posture of the motion is described by a graph whose vertices are characterized by differential information and some edges are associated with distance constraints. Differential information is used to preserve the spatial relationships between markers while distance constraints enable to preserve the length of some edges.
In this document, we propose to study other representations of the motion through a set of spatial relationships. Thus, we propose two approaches~: the first one considers the motion in the metric space and the second one characterizes each posture by a differential representation using the discrete Laplacian operator.
Representation of the motion based on the Laplacian expression of a 3D+t graph: the set of connected graphs given by the skeleton over time. Through this Laplacian representation of the motion, we propose an application which allows an easy and interactive editing, correction or retargeting of a motion.
This project advocates the idea that the metric space (i.e. set of joint-to-joint distances of the skeleton), though bigger, can yield a powerful representation of the motion. In order to validate this idea, comparisons with rotation and position based representations are experimented in a motion retrieval task performed with the HDM05 motion database. The final resulting retrieval algorithm is very fast with satisfying performances.
This study proposes a general method for driving kinematics by distances and more specifically for controlling kinematically articulated systems. Unlike traditional approaches, the problem is addressed in the metric space using distances belonging to points of the skeleton and to the environment. After defining kinematic control through a distance-based formalization, we propose an optimization method for solving classic issues such as motion adaptation and inverse kinematics. The originality of the method lies in the possibility to introduce distance constraints with priorities. The approach is validated by a large variety of experiments in the field of motion control of articulated figures, and compared to other approaches by means of stability, convergence and performance issues.
In this paper we present a multichannel animation system for producing utterances signed in French Sign Language (LSF) by a virtual character. The main challenges of such a system are simultaneously capturing data for the entire body, including the movements of the torso, hands, and face, and developing a data-driven animation engine that takes into account the expressive characteristics of signed languages. Our approach consists of decomposing motion along different channels, representing the body parts that correspond to the linguistic components of signed languages. We show the ability of this animation system to create novel utterances in LSF, and present an evaluation by target users which highlights the importance of the respective body parts in the production of signs. We validate our framework by testing the believability and intelligibility of our virtual signer.
Face à l'engouement (plusieurs milliers de participants) pour cette mythique course valaisanne, la question des ralentissements dans certaines portions étroites vient la problématique de la gestion des flux humain. Nous avons développé un simulateur de foule appliqué au parcours de Sierre-Zinal (longueur, dénivelé, largeur des chemins) et l’avons appliqué à quelques 5000 agents répartis selon des caractéristiques de performance représentative de la population amateur. Suite aux résultats concluants (quelques minutes d’erreur, 10mn par participant), les organisateurs ont modifié la configuration de la course.
Implemented in C++ -- Projet réalisé avec Guillaume Maire et Valentin Genoud.