410 likes | 592 Views
3D Models for Face Image and Video Processing. Gábor Szirtes ELTE Dept of Information Systems Neural Information Processing Group Lőrincz-lab. Content. Short Introduction Motivations Pile of concepts ( framework ?) Future applications Ongoing projects and immature ideas What next?.
E N D
3D Models for Face Image and Video Processing Gábor Szirtes ELTE Dept of Information Systems Neural Information Processing Group Lőrincz-lab SSIP 2002 Budapest, Hungary
Content • Short Introduction • Motivations • Pile of concepts(framework?) • Future applications • Ongoing projects and immature ideas • What next? SSIP 2002 Budapest, Hungary
A few words about our group • Since 1999 • Within the Information Systems Department, ELTE • 5 PhD students and ~20 grad students • Mainly biologically motivated projects • RL, ICA, machine learning, facial expressions, dynamical systems, image processing SSIP 2002 Budapest, Hungary
Motivations • Being in quest of the ‘Holy Grail’: intelligence • One working example: our brain • Evolutionary concepts, need for adaptation • Perception and Action SSIP 2002 Budapest, Hungary
Conceptual framework I. ? SYSTEM Perception Action ENVIRONMENT Noisy, stochastic, evolving SSIP 2002 Budapest, Hungary
Conceptual framework II. Central hypothesis • Internal representation (encoded signals from the environment and the system’s state) • Reconstruction SSIP 2002 Budapest, Hungary
Perception • Active • Not simply feed-forward • Feed-back modulated and controlled • Modular • Component based • Adaptive, ‘plastic’ SSIP 2002 Budapest, Hungary
Perception II. • Active: it is not a passive signal detection process. We need to `foresee` and anticipate the expected changes (prediction). • Influenced by higher order modulation (e.g. FOA, focus of attention, conscious and unconscious perception) SSIP 2002 Budapest, Hungary
Perception III. • Several stages of processing • Not purely hierarchical (feed-back) • Distributed, parallel ways, strong interplay Modularity SSIP 2002 Budapest, Hungary
Perception IV. Components: meaningful (?) building blocks SSIP 2002 Budapest, Hungary
Perception V. This is what we have seen before? SSIP 2002 Budapest, Hungary
Parts… Drawings of 4 year old healthy children SSIP 2002 Budapest, Hungary
…and the whole Drawing of a 3 and a half year old child with autism SSIP 2002 Budapest, Hungary
Beyond the theory Recognition of faces and facial expressions • Twofold goals: • Understand perception • Help develop applications for Human-Computer Interaction Psychiatric analysis and treatment … SSIP 2002 Budapest, Hungary
Database In collaboration with the Psychiatric Clinic of SOTE (Simon-lab) SSIP 2002 Budapest, Hungary
A few examples of segmented images Happiness Disgust SSIP 2002 Budapest, Hungary
The architecture ACTION ? RL container M* M** M1 M2 M3 FACES SSIP 2002 Budapest, Hungary
Modules for recognition of faces • Finding heads: Skin detection • Tracking: particle filtering • Segmentation • 3D model based transformation • Identification, recognition or analysis • (back-transformation) SSIP 2002 Budapest, Hungary
Module 1 Face location (fitting) • Many heuristics are possible • One particular choice skin-detector SSIP 2002 Budapest, Hungary
Skin detector b g Skin color cluster learned by MLP r SSIP 2002 Budapest, Hungary
Module 2+ Particle filtering Segmentation Tracking CONDENSATION (Conditional Density Propagation ) (Isard and Blake, 1998) SSIP 2002 Budapest, Hungary
Segmentation Image based Feature based procedures • Two approaches: • approximating contours with splines or snakes (too many degrees of freedom) • Template based A simple template SSIP 2002 Budapest, Hungary
Segmentation II. More sophisticated manually tuned template Arbitrary spine directions (with positive-negative weights) SSIP 2002 Budapest, Hungary
Segmentation III. Many concurrent candidates SSIP 2002 Budapest, Hungary
Segmentation IV. Head-shoulder template for better fitting SSIP 2002 Budapest, Hungary
Particle filtering in action! Initialization made by hand SSIP 2002 Budapest, Hungary
Well, there is no perfect method… Sometimes even the best choice is far from the face to be tracked… SSIP 2002 Budapest, Hungary
Tracking of fast motion against a cluttered background From http://www.robots.ox.ac.uk/~misard/condensation.html SSIP 2002 Budapest, Hungary
CONDENSATION Keywords: general, multi-modal densities, sampling, Discrete -continuous Markovian SSIP 2002 Budapest, Hungary
Module 3 (off the stream) Facial expression (display) recognition SSIP 2002 Budapest, Hungary
Facial expression recognition surprise HMM winner: surprise HMM on segmented image sequences Reconstruction error HMM emission SSIP 2002 Budapest, Hungary
Module 4 3D face model Extension of the CANDIDE (Rydfalk,1987) model Compatible with FACS (Ekman and Friesen, 1977) Candide 3 (developed for MPEG4 standard) SSIP 2002 Budapest, Hungary
How to use the model? Target (synthetic) face Searching SSIP 2002 Budapest, Hungary
Such a big space! • Reconstruction error based optimization problem • Too many local minima • Global optimum finding procedure: STAGE (Boyan, 1998) SSIP 2002 Budapest, Hungary
STAGE • Algorithm for finding the global optimum • Function approximator • learns an evaluation function that predicts the outcome of a local search • Experience: it is able to explore the global structure Let us find the minimum of F(x)=(|x|-10)cos(2x) SSIP 2002 Budapest, Hungary
STAGE II • It can be combined with any local search method (hillclimbing,WALKSAT,…) • It works on both the objective and the evaluation function at two stages • Smart restart by a better prediction • Real-valued (compared to GA) • Easy to implement SSIP 2002 Budapest, Hungary
What we have got so far? • A few working modules • Working RL architectures • Working combination in an other problem domain: Internet search • …and research is focused on how to link all of our concepts. SSIP 2002 Budapest, Hungary
What next? • Many avenues • Ongoing projects with psychiatrists: trajectory analysis with cliplets, transient expressions, depression quantification… • Distance learning • Human Computer Interaction • Virtual reality SSIP 2002 Budapest, Hungary
Infos about our research activityhttp://people.inf.elte.hu/lorincz/ SSIP 2002 Budapest, Hungary
Thanks for your attention (and patience)! SSIP 2002 Budapest, Hungary