Versão
Portuguesa João Maciel: Final Project
Title Robot Sketch Synthesis (document in Portuguese)(.dvi version)( short paper)
Coordinators
Subject
I'm develloping a system that synthetises human face images,
based on natural language descriptions and, possibly, on an interaction
scheme between the machine and the user. Systems like this are already
used by Police laboratories, but they present sketches built by arranging
standard pieces of face images (eyes, noses, ...), resulting images with
unreal look. This project has also the porpose of studying the mathematic
fundamentals of the approach used, wich can be applyed to many other problems.
This approach consists on considering the images as points in a multidimensional
space (dimension equals the number of pixels). Using a face
base with high number of examples, one can sample the subspace of human
face images. The research efford has been applied to find efficient face
subspace representations, wich allows us to interpolate between different
examples, inside the face subspace (morphing), using a minimum possible
number of parameters: 
I'm also trying to identify the structure of this subspace and relate it with the holistic properties of the images (expression and other parameters used in high level descriptions). This approach can be generalized to other applications – automatic security systems, based on face recognition; image coding breakthroughs with application in teleconferencing and image databases; computer graphics, animation and virtual reality – and other research areas – classification in general; associative memories; statistics and exploratory multidimensional data analysis. The greatest difficulties come mainly from the subjectivness of the descriptions and difficulty on visualising felt by the user, resulting on descriptions poor on information that can lead to a good reconstruction; this forces the interactive step. Furthermore it is necessary to deal with the complex and related effects of variations in pose, illumination, expression, background ... The hope is the fact that these effects seem not to affect the human ability to recognize faces; this suggests that it might be some approach or coding scheme that reduces the problem to triviality!!
This page is still under construction
and is maintained by Ana Saianda
Last update: June 1997