Minutes from BUT taken during the meeting

= MOBIO IDIAP 8-9.1.2009 minutes =

= Annual report [Francois Foglia] =
* everything is on the SVN in Latex.
* should be able to  compile the Latex file.
* we'll complete the file - WP. We need to complete chapter 3 and chapter 6.
=> provide the  COST TABLE to SYLVA !!!
* new: filling the form C on-line ! Project number is '214324', the code is on the slides by Francois.
* we won't need to send signed form C unless the EC application does not work.

== Audit certificate ==
* only if cummulated financial contribution exceed 375000 EUR. Not this year.

= Overview of the project in 2008 [Sebastien] =
* everything on good track :-)
* should reacit to mails asking contribs to deliverables.

= Overview of the DB [Chris McCool] =
* need to make people happy, if we have drop-outs, we have a problem !
* there will be a feed-back form (5-10 min to cmplete).
=> give the form to the people collecting the data.
... we'll probably to translate this to Czech ... and accept answers in Czech.
* can also capture data in the mean-time if we want.

= Publications [HOnza] =
* asking for updates !
=> send info on Marcel's paper at Goa and ICASSP 2009 to MOBIO list.

= D3.1 (SW) info [UMAN] =
* all in place, except LIA.
* will put this in some nice format of archive.

= D6.2 [Giorgio] =
* submitted, will be discussed tomorrow.

= D2.1 [Giorgio] =
* remarks from the reviewers - 'trust and threat issues'
* partitioned the secutiry into: data confidentiality, data integrity and data availability
* added discussion of this to the use cases. good.
* the deliverable is restricted but we want to distribute the scenarios to CoI members.
* EPM will check if it is OK to send this to them.

== ThinkTrust project interest ==
* they want our use-cases, at the end of the year, they can make them public by the end of the year.
* Giorgio will check.

= Progress reports =
* BUT was OK, pretty much the same as from IDIAP.
* Comments: should probably train the UBM on BANCA or at least train the channel comp. put this into the D3.3 (or D3.4 ?) deliverable !
* Manchester - Random Markov fields for face ID, will play with AdaBoost !
* EPM ...
* SUrrey:
** MBGC - Multi-Biometry Grand Challenge - also 1 image to enroll and video to test! the organizer is NIST.
** working also with someone from sociology on the useability.
** ICB 2009 competition - 4-5 groups participating, some with 2 systems, IDIAP with 11 systems ;-) controlled vs. degraded scenario. Will be evaluated also by the EPC (expected performance curve) - should be close to 'ultimate' results. But will publish also DETs.

= WP4 [Norman] =
== score level fusion ==
* score level fusion - fixed interval or after sufficient data was processed and we have some idea of reliability ?
* Surrey have 14 quality measures.
* Chris: why not use these features directly to classify ?
* Norman and Joseph: features not necesarily good to discriminte people (lighting...) but can give a clue on how good the decision is.

== feature level fusion ==
* maybe giving something features to Norman ?

== model level ? ==
* Sebastien wanted this but is not happy with it, too difficult.
* Joseph: lip dynamics as biometrics.
* HOnza: can help face people by phone recognition, maybe for some phone conditioning ? said this to guys, will tell if interested.

== Speech quality measures ? ==
* are we interested in ? hm hm.

== How we could contribute ? ==
see slides for proposals.

== IDIAP propopsal ==
* bi-modal auth and model adaptation, according to some Campbell's paper, polynomial expansion of the 2 streams, then something.
* then something else ...

== UMAN proposal ==
??

== LIA proposal ==
* integrating unsupervised progressive adaptation into FA framework.
* studying shift in scores for progressive adaptation (the amount of data increases for the prog. adaptation which can play on scores).
* have 2 Gaussians trained for client and impostor scores, can derive some confidence out of this. Do not do any decision when not sure. On NIST 2005 data:
** have a decrease from EER 3.84% to 0.03% and 50%  of utterances non-treated.
** have a decrease from EER 3.84% to 1% and 30%  of utterances non-treated.

== UNIS proposal ==
* adaptive vs. non-adaptive fusion, "advanced adaptive fusion" - aka "co-training" audio and video help each other.

= D4.1 planning =
* fusion of audio and video classifiers.
* scores will be made publicly available - OK.

= D4.5 and D4.6 - ADAPTIVE SYSTEMS =
* LIA will do this for speech - baseline, not optimize, nothing.
* can we use BANCA for this ? YES, needs to use 1 session for training, has total of 12 sessions, so can use 1 for training and 11 for progressive adaptation.
* will define 3 sets on BANCA
** all client accesses
** all impostor accesses
** mixed
** finally will have 50% 50% mix of clients and impostors.
... as Driss claims the amounts are very important (people got nice results for NIST 2005 but terrible for 2006!)
* the final results should be comparable to the standard system.
* can also have the oracle performance - "perfect adaptation system".
* LIA will provide a system to do incremental stuff for speech - good.
* delivered in May. 


= TASKS =
=> give the form to the people collecting the data.

=> finish FAST the report on publications and patents, send to Valerie

=> send info on Marcel's paper at Goa and ICASSP 2009 to MOBIO list.

=> do not forget to put participation in NIST SRE 2008 in the annual report !

=> check with Paja / Lukas whether we're intersted in quality measures


BIB TEX is accepted !!!!


= WP3 planning =
* if you have face points (eys, mount, nose) UMAAN want to transform face accoring this points to then neutral face, and will provide this faces - provide executable (in mid march) to do this

* IDAP - new methods for FD (fase detection) - boosted Haar-LBP
* OULU - imporoved LBP-SVN for FD, new features+detection scheme+learning method, Weber law based system
* UMAN - improoved FL (fase localization)
* UNIS - improoved LBP based FV (fase verif)
* IDIAP - gmm-subband system for FV
* VAD is enought good,
* Speaker Ver - send scores on BANCA to LIA to the can plot both DET plots to one graph. cominicate with LIA to we get the same results on the one scores!!!
** LIA - noise compensation dealing with target data, trying to incorporating it to the model domain
** BUT - full vactor analysis
** next time we should make theoretical presentation about FA, what is FA about

* evaluation - report is m22 (end 2009) evaluate on BANCA and part of MOBIO (probably), use svn as before.


= WP5 planning =
* shift deliverable to november, but see progres from jun
* some parametr to tune the scalability of the system (one with 2048 gauss,....,128 gauss)
* where we can reduce the computation? 10best, tree clustering of gaussions
* model complexity vs selection and quantization
* need confidence score
* BANCA, for the report on MOBIO
* what are the system which should be deliverd? who, what, when
** what will be on the server what on the client? we should discuss this with LIA (Driss) are we going to deliver client based application? or only server based? what is written in proposal?
** no one is going to implement things in mobile, we should address the fixpoint operation, Sebastian is going to report for his system what happen when you go from floating point to fix point. Who else? talk to driss - he is responsible for deliverable.
** Driss send email with instructions till end of january and we should replay in two weeks - deadline for us mid february -> resulst from this is who does what and when -> roadmap for us.
** we run the evaluation of our scalable system, results are:
*** mobio results file - performence
*** time delay - we define
*** confidence measure for each system


= Anual Report =
* svn
* duty as WP manager be sure that all partners fill it up

= WP6 =
== D6.2 ==
* D6.2 take a look at EMStreamFilter.h
* they will send the real examle by march
* input and output in each frame
* for audio - ech 10ms we receive audio data  (10ms of audio)
* the system can sudenly change sampling rate (more users, small bandwith, switch from 16K to 8k ) we have to deal with this by ourselve, => system will work on 8kHz and will be there block dor resampling - they have some filter for that, but this is too complicated, because, if we put in front of our filter resampling filtr then at the end there have to be filter for upsampling ....
* wants windows platform
* mobile devices do not have swap disk - DO NOT EXCEED RAM - we have 32 to 64MB max.

== D6.3 ==
* identify problems which should be considerd, like third party libraries, what kind of functionality we are using out of it (blas, ....)
* user interface, how does it sould like? will go to 6.2 mailing list
* until I receive 5 sec of audio my output will be -100000,
* deadline june !!
* end of January *.cpp *.h, during February the rest


= WP7 =
* competitions
** ICB take from anual report - norman will not send more info, if we do not ask for it
* also inveted talk (Joseph Kitler mentioned that)

= Database =
* more realistic conditions
* impostor attack
** personal information - name, adress, other data
** hangig photo in front of the camare while talking
** different people? have to sign consent form!
** how many impostor attacs?
** you can do it manually by switching audio and video channel between recordings
** simply do not do impostor attack , but normal impostors from all other people
** each site sould write what thay want from impostor attack by end of Feb.
** other?
* outdoor data
** how many data we need? do we need it?
** only outside from office
** try to record some outside or outside of office and see difficulties.
* Recordings
** Kamil write to Chriss what problems he had during recordings
** What he knows from talking with people who record database?