Computer Vision Group
September 1992


The Computer Vision group (currently 9 persons) is part  of the  AI and Vision
research team,  which  comprises  approximately  16  members.    The Computing
Science Center  of the  University of  Geneva hosts  approximately 75 members,
including 12 professors and 5 to 10 post-docs.

The Computer  Vision group  is headed  by Prof.   Thierry  Pun, the Artificial
Intelligence group by Prof.  Christian Pellegrini.


Main interests in computer vision

We pursue two series of objectives.  First, on the theoretical side, we aim at
a better understanding, modelling and usage of the  knowledge concerning human
visual perception.    We  think  that  this  should  lead to  new concepts and
results, of value for pre-attentive vision as well as for higher levels of the
system.  We are currently developing a general purpose vision  system based on
primal indexing ("Geneva Vision System").

Second, we wish to apply our  work to  practical problems.   We  have tried to
identify a  sector  with  practical  importance,  as   well  as  theoretically
non-trivial.  This  is  the  case  with  agricultural  robotics,  or  agrotics
("Potato Operation").  We  have also  developed an  integrated environment for
research and teaching in image analysis ("LaboImage").


Geneva Vision System

Our project  aims  to  realize  an  artificial  vision   system;  the  central
recognition mechanism relies upon asynchronous identification in  the scene of
tridimensional objects  whose  models are  a priori  known.   We are currently
concentrating our  efforts  on  the  primary  access  problem.   The following
subjects are being investigated (from "low" to "high" level):

    -	features extraction using the concepts of asynchronicity and coherence;
    -	context and shape inference using non-accidental properties, in a
	world of geons;
    -	active vision and motion;
    -	grouping using an electromagnetic approach;
    -	extraction of spectral and intrinsic characteristics;
    -	definition and implementation of a focus of attention mechanism;
    -	indexing using an asynchronous network of discriminative knowledge 
	bases;
    -	learning of visual concepts.

At the same time, a software architecture is being developed.


Potato Operation

Every year billions of potatoes are used  as seeds,  and therefore  need to be
sampled for  possible  viral  infections (potato  viruses).   This sampling is
currently done manually, by locating large germs near  the top  of the potato,
then digging  a drill  into them.   The  pulp extracted  by the  drill is then
analyzed using an Elisa test.

The goal of the Potato Operation is to automatize  the sampling  of the potato
pulp.  The project involves two aspects:

    -	image analysis and computer vision, in order to find the position where
	the drill has to sample the pulp. This necessitates dynamical analysis
	and active vision. Multiple views of the potato are necessary to 
	discriminate between germs, scars and other defects, as well as to 
	correctly locate the drilling position;
    -	robotics, in order to design a system with a camera, an arm, a drill, 
	and pulp containers, able to grasp a potato; orient it with respect to 
	the drill; drive the drill into the potato; deposit the pulp sample in 
	a container.

This project  is  being  conducted  in  collaboration  with  the Swiss Federal
Agricultural Station in Changins,  and the  Institute of  Microtechnics at the
Swiss Federal Institute of Technology.


LaboImage

We are distributing (free) LaboImage, an image analysis software  that we have
developed during the past 5 years.  It is used for teaching and  research.  It
has a  menu-driven  interface, an  extensive set  of image processing/analysis
capabilities (filtering,  transforms,  segmentation,   binary  and  grey-level
morphology, etc.),  some  special  purpose  tools (particles  counting, 1D gel
analysis, etc.).  It  is written  in C  (approximately 80'000  lines of code).
The latest version runs with Motif-X11.


Equipment

General purpose hardware of the Computing Science Center:

    -	PCs, MacIntoshes;
    -	Connection Machine CM2a-8S-FS (8'192 processors, 256MB memory);
    -	Volvox TS 408-204 (32 transputers with 4MB memory each);
    -	Hewlett-Packard: 9000/835;
    -	Next;
    -	Silicon Graphics: IRIS PC;
    -	Suns: Sun 3 series, Sun 4 series.

The AI and Vision people mostly work on Sun hardware.   There is approximately
one workstation per researcher.

Specialized hardware:
    -	Dunn Microcolor 638 HSC, image hardcopy on films;
    -	video camera MINTRON MTV-1003CB, 512x512x8 + Occulus board;
    -	Eikoniks 1412 camera, 4096x4096x12 + VME board (on Sun 3/160).


Contact

Prof. Thierry Pun, Computer Vision Group
Computing Science Center, Uni-Dufour
24, rue du General Dufour, CH-1207 Geneva 4, SWITZERLAND
Phone: +41(22) 705 7627; secr: x7660; fax: x7780; e-mail: pun@cui.unige.ch
