48
Deep learning and egocentric vision 1 Pe3a Radeva & Mariella Dimiccoli www.cvc.uab.es/~pe. a Barcelona Perceptual Compu.ng Laboratory (BCNPCL), Universitat de Barcelona & Computer Vision Center

CAIP 2015 Introduction on Visual Lifelogging I part

  • Upload
    ledien

  • View
    226

  • Download
    2

Embed Size (px)

Citation preview

Page 1: CAIP 2015 Introduction on Visual Lifelogging I part

Deep  learning  and  egocentric  vision  

1  

Pe3a  Radeva  &  Mariella  Dimiccoli    

www.cvc.uab.es/~pe.a    

Barcelona  Perceptual  Compu.ng  Laboratory  (BCNPCL),  Universitat  de  Barcelona  &      

Computer  Vision  Center  

Page 2: CAIP 2015 Introduction on Visual Lifelogging I part

Index  

ì  1.  Egocentric  Vision  and  Lifelogging  

ì  2.  Deep  learning  

ì  3.  Temporal  egocentric  segmenta.on  for  event  extrac.on  

ì  4.  Social  analysis  from  egocentric  images  

2  

Page 3: CAIP 2015 Introduction on Visual Lifelogging I part

Index  

ì  1.  Egocentric  Vision  and  Lifelogging  

ì  2.  Deep  learning  

ì  3.  Temporal  egocentric  segmenta.on  for  event  extrac.on  

ì  4.  Social  analysis  from  egocentric  images  

3  

Page 4: CAIP 2015 Introduction on Visual Lifelogging I part

Wearable  cameras  and  the  life-­‐logging  trend  

IMEC,  University  of  Eindhoven,  March,  2015   4  

Shipments  of  wearable  compu3ng  devices  worldwide  by  category  from  2013  to  2015  (in  millions)  

Page 5: CAIP 2015 Introduction on Visual Lifelogging I part

Technology  is  “running”!  

Evolu.on  of  life-­‐logging  apparatus,  including  wearable  computer,  camera,  and  viewfinder  with  wireless  Internet  connec.on.  Early  apparatus  used  separate  transmiVng  and  receiving  antennas.  Later  apparatus  evolved  toward  the  appearance  of  ordinary  eyeglasses  in  the  late  1980s  and  early  1990s.  

5  “Quan.fied  Self  &  life-­‐logging  MeetsInternet  of  Things  (IOT)”,    Mazzlan  Abbas.  

Page 6: CAIP 2015 Introduction on Visual Lifelogging I part

Visual  life-­‐logging  Defini3on:  Visual  life-­‐logging  consists  of  acquiring  images  related  to  an  individual  through  a  wearable  camera.  

Benefits:  

ì  A  digital  memory  of  people  you  met,  conversa3ons  you  had,  places  you  visited,  and  events  you  par.cipated  in.    ì  This  memory  would  be  searchable,  retrievable,  and  shareable.  

ì  A  14/7/365  monitoring  of  daily  ac3vi3es.    ì  This  data  could  serve  as  a  warning  system  and  also  as  a  personal  base  upon  which  to  diagnosis  illness  and  to  prescribe  medicines.  

ì  A  way  of  organizing,  shaping,  and  “reading”  your  own  life.  ì  A  complete  archive  of  your  work  and  play,  and  your  work  habits.  Deep  compara.ve  analysis  of  your  ac.vi.es  could  assist  your  

produc.vity,  crea.vity,  and  consump.vity.  

ì  To  the  degree  this  life-­‐log  is  shared,  this  archive  of  informa.on  can  be  leveraged  to  help  others  work,  amplify  social  interac3ons,  and  in  the  biological  realm,  shared  medical  logs  could  rapidly  advance  medicine  discoveries.  

6  

Page 7: CAIP 2015 Introduction on Visual Lifelogging I part

Will  life-­‐logging  and  internet  of  things  help  know  us  beber?  

7  

Don’t  you  love  to  know…    •  Where  you’re  going?!  Who  you’ve  interacted  with?!  

•  How  long  you’ve  spoken  to  friends?!  The  affinity  to  connec.ons?!  

•  How  long  it  takes  to  get  to  work?!  

•  The  tone  of  your  messages?!  The  amount  you  text,  tweet,  or  update?!  

•  How  much  exercise  you’re  geVng?!  

•  How  much  you  get  distracted?!  Where  is  your  .me  most  spent?  

Page 8: CAIP 2015 Introduction on Visual Lifelogging I part

Will  life-­‐logging  and  internet  of  things  help  know  us  beber?  

8  

Life-­‐logging  everything  you  see,  do,  feel,  speak,  experience  and  hear  is  almost  here.    

Wearable  camera  can  provide  many  benefits,  such  as  assis.ve  technologies  to  help  people:    

•  see  beber,    •  store  and  remember  beber,    •  work  beber,    •  func.on  beber,    •  remember  and  recognize  names  and  faces…  

Page 9: CAIP 2015 Introduction on Visual Lifelogging I part

AMDO'2014   9  

The  three  trends  of  the  mainstream  Lifelogging  •    Lifelogging  is  keeping  everything  by  

using  technology  to  capture  events,  documents  and  experiences  and  keep  them  in  a  chronological  .meline.  

     Life-­‐streaming  (social  networking)  •    Life-­‐streaming  is  sharing  details  of  your  life  

with  other  people.      Self-­‐quan3fied  (health  &  wellfare)  •    Monitors  your  biological  status  that  you  

capture  but  for  the  purpose  of  improving  and  sharing  

     

Page 10: CAIP 2015 Introduction on Visual Lifelogging I part

Social  networks  &  lifelogging  ì  Facebook  originally  began  as  a  college  networking  website  in  2003  and  has  expanded  beyond  anyone  

could  imagine.    ì  Facebook  has  more  than  350  million  ac.ve  users  and  more  than  2.5  billion  photos  are  uploaded  every  month.  

ì  Flickr.com  is  an  image  and  video  hos.ng  website  launched  in  2004  by  a  company  called  Ludicorp.    ì  Flickr  offers  its  users  a  comprehensive  photo  organisa.on  and  tagging,  which  makes  it  the  web‟s  Number  1  

website  for  bloggers  to  share  their  photos.    

ì  Flickr  has  around  44  million  users.    

AMDO'2014   10  

“To  Log  or  Not  to  Log?  Privacy  Risks  and  Solu.ons  for  Lifelogging  and  Con.nuous  Ac.vity  Sharing  Applica.ons”,  Blaine  Price,  ENISA,  Centre  for  Research  in  Compu.ng.  

•  Youtube  is  the  most  popular  video  sharing  site  that  lets  anyone  upload  short  videos  for  private  or  public  viewing.  •  Founded  in  2005  by  Chad  Hurley,  Steve  Chen  and  

Jawed  Karim,  Youtube  is  one  of  the  Top-­‐5  most  visited  websites  in  the  world  with  more  than  5  billion  videos  uploaded  monthly.    

 

Page 11: CAIP 2015 Introduction on Visual Lifelogging I part

Social  networks  &  lifelogging  

ì  Twiaer.com  is  a  social  networking  website  created  in  2006,  offering  a  micro-­‐blogging  service  that  lets  users  send  and  read  short  text  messages  known  as  tweets  up  to  140  characters  in  length.    

ì  Twiber  is  becoming  more  popular  and  as  a  result,  more  than  3  million  tweets  are  published  daily.    

AMDO'2014   11  

•  Instagram  community  has  grown  to  more  than  200  million  Instagrammers  capturing  and  sharing  their  lives  every  month.  

•  As  Instagram  exceed  20  billion  photos  shared  on  Instagram  to  date,  we  look  back  in  wonder  at  the  beauty  and  importance  of  everything  this  community  has  created.  

 

Page 12: CAIP 2015 Introduction on Visual Lifelogging I part

Health,  wellfare  &  lifelogging  

What  is  Quan.fied  Self  (QS)?    The  con.nuous  tracking  of  various  aspects  of  our  physical  bodies:  •  how  many  calories  we  burn;      •  how  many  hours  a  week  we  spend  commu.ng  or  siVng;...  •  our  body  fat  percentage;  •  how  many  steps  we  take  in  a  day;  •  how  long  we  sleep;…  

AMDO'2014   12  

Source:  Swan,  M.  Sensor  Mania!  The  Internet  of  Things,  Objec.ve  Metrics,  and  the  Quan.fied  Self  2.0.  J  Sens  Actuator  Netw  2012.  

How  many  self-­‐trackers/self-­‐quan3fiers  are  there  in  the  audience?  

Increasingly  con3nuous  and  automated  data  collec3on  for:  Physical  Ac3vi3es  •  Miles,  steps,  calories,  repe..ons,  sets,  METs1    Diet  and  Nutri3on  •   Calories  consumed,  carbs,  fat,  protein,  specific  ingredients,  glycemic  index,  sa.ety,  

por.ons,  supplement  doses,  tas.ness,  cost,  loca.on    Psychological,  Mental,  and  Cogni3ve  States  and  Traits  •   Mood,  happiness,  irrita.on,  emo.on,  anxiety,  esteem,  depression,  confidence    IQ,  alertness,  focus,  selec.ve/sustained/divided  aben.on,  reac.on,  memory,  verbal  fluency,  pa.ence,  crea.vity,  reasoning,  psychomotor  vigilance    Environmental  Variables  •   Loca.on,  architecture,  weather,  noise,  pollu.on,  cluber,  light,  season    Situa3onal  Variables  •   Context,  situa.on,  gra.fica.on  of  situa.on,  .me  of  day,  day  of  week    Social  Variables  •   Influence,  trust,  charisma,  karma,  current  role/status  in  the  group  or  social  network  

Page 13: CAIP 2015 Introduction on Visual Lifelogging I part

Living  “digitally”  

Life  signals:  The  marketplace  is  soon  to  be  flooded  with  inexpensive  sensors  (smart  phones  already  integrate  cameras,  microphones,  GPS,  and  accelerometers):  

ì   collect  a  wide  variety  of  data   in  digital  form—e.g.,  con3nuous  video  and  sound  of  one’s  daily  ac3vi3es,  sport  performance,  body  weight,  or  sleep  paaerns—  

ì automa.cally  geoloca.ng  and  uploading  personal  data  to  websites  ,  

ì providing  sta3s3cal  and  visualiza3on  tools  for  forecas.ng  and  trend  analysis.    

 

 

 

 

 

 

 

 

 

 

 

 

 

Con3nuous  and  complete  life  signals:  If  one  factors  in  the  gradual  elimina.on  of  paper  in  favor  of  digital  forms  for  commercial  transac.ons,  communica.on,  documenta.on,  etc.,  and  the  con.nually  plumme.ng  costs  of  digital  storage,    

ì   the  picture  of  a  world  where    “everyone  is  on  the  record  all  the  3me”  does  not  seem  far-­‐fetched.    

 

Digital  &   on   the   cloud:   Furthermore,     for   the   first   .me   in   history,   these   different   types   of   records  —   images,     quan.ta.ve   data,   audio  recordings,  wriben  documents,  etc.  —  are  available  for  transmission,  storage,   indexa3on,  analysis,  retrieval,  and  visualiza3on  in  a  single  media.  

AMDO'2014   13  

Page 14: CAIP 2015 Introduction on Visual Lifelogging I part

Do  you  know  that…?  

hbp://ideas.ted.com/2014/11/17/the-­‐economic-­‐impact-­‐of-­‐bad-­‐mee.ngs/  

Page 15: CAIP 2015 Introduction on Visual Lifelogging I part

University  of  Groningen,  November,  2014   15  

Do  you  know  that…?  

Page 16: CAIP 2015 Introduction on Visual Lifelogging I part

University  of  Groningen,  November,  2014  

hbp://ideas.ted.com/2014/11/17/the-­‐economic-­‐impact-­‐of-­‐bad-­‐mee.ngs/  

Do  you  know  that…?  

Page 17: CAIP 2015 Introduction on Visual Lifelogging I part

Mee.ngs  and  us  

ì  Is  our  impression  about  our  .me  spent  correct?  

ì  How  to  extract  a  summary  of  the  mee.ng  event  in  order  to  display  it  and  op.mize  our  ac.vi.es?    

ì  What  about  recording  all  our  day,  day  by  day?    

CAIP,  Malta,  September,  2015   17  

Page 18: CAIP 2015 Introduction on Visual Lifelogging I part

Is  it  feasible  to  record  everything  that  happens  in  a  person’s  life?  

ì  In  1970,  a  disk  to  store  20  MB  was  the  size  of  a  washing  machine  and  costed  20.000$.  

ì  Today  a  TB  (one  trillion  bytes)  costs  a  100$  and  is  the  size  of  a  paperback  book.    

ì  By  2020  a  TB  will  cost  the  same  as  a  good  cup  of  coffee  and  will  probably  be  in  your  cell  phone.    

ì  100$    will  then  buy  around  250  TB  of  storage,  enough  to  hold  tens  of  thousands  of  hours  of  video  and  tens  of  millions  of  photographs.    ì  This  should  sa.sfy  most  life-­‐loggers’    recording  needs  for  an  en3re  life.  

ì  In  fact,  digital  storage  capacity  is  increasing  faster  than  our  ability  to  pull  informa.on  back  out.      ì  From  2000  it  became  trivial  and  cheap  to  sock  away  tremendous  piles  of  data.    

 

“Total:  How  the  E-­‐memory  revolu.on  will  change  everything”,  Gordon  Bell  and  Jim  Gemmel,  2009,  Dubon,  Penguin  Group.    

18  

The  Moore’s  Law  (1965):    “Transistor  density  that  can  be  etched  onto  the  silicon  wafer  of  a  microchip  doubles  every  two  years”.    

Page 19: CAIP 2015 Introduction on Visual Lifelogging I part

Is  it  feasible  to  record  everything  that  happens  in  a  person’s  life?  

“Total:  How  the  E-­‐memory  revolu.on  will  change  everything”,  Gordon  Bell  and  Jim  Gemmel,  2009,  Dubon,  Penguin  Group.    

19  

The  Moore’s  Law  (1965):    “Transistor  density  that  can  be  etched  onto  the  silicon  wafer  of  a  microchip  doubles  every  two  years”.    

The  hard  part  is  no  longer  deciding  what  to  hold  on  to,  but  how  to  efficiently  organize  it,  sort  it,  access  it,  and  find  paaerns  and  meaning  in  it.        

This  is  a  primary  challenge  for  the  engineers  that  will  fully  unleash  the  power  of  Total  Recall.    

Page 20: CAIP 2015 Introduction on Visual Lifelogging I part

Data  management  

CAIP,  Malta,  September,  2015   20  

The  really  big  issue  here  is  that  you  might,  individually,  not  worry  about  recording  and    even  publishing  details  of  your  personal  life.  

 • But  you  are  publishing  your  friends,  family  and  business  contacts  details  at  the  same  3me.    

Page 21: CAIP 2015 Introduction on Visual Lifelogging I part

Ethical  guidelines  for  wearable  cameras  

ì  Anonimity  and  confiden3ality:  Researchers  coding  image  data  should:  ì  not  discuss  the  content  with  anyone  outside  of  the  team,  ì  not  iden.fy  anyone  they  recognize  in  the  images,    

ì  be  aware  of  how  sensi.ve  the  data  are.  

ì  Data  encryp3on:  Conxden.ality  can  be  protected  by  conxguring  devices  and  using  specialist  viewing  soyware  to  make  the  images  accessible  only  to  the  research  team  (lost  devices).  ì  Devices  should  be  configured  so  that  data  can  only  be  retrieved  by  the  research  team.  It  

should  be  impossible  for  par.cipants  or  third  par.es  who  find  devices  to  access  the  images.  

ì  Data  storage:  Collected  images  should  be  stored  securely  and  password-­‐protected,  according  to  na.onal  regula.ons.  

 21  

Page 22: CAIP 2015 Introduction on Visual Lifelogging I part

Understanding  user  privacy  requirements  and  risks  from  emerging  technologies  

ì  People  are  bad  at:  1.  understanding  the  future  value  of  revealing  private  informa.on  

today,  

2.  understanding  the  risks  from  technology  they  have  not  yet  used  or  heard  of.  

22  

Wearable  cameras  can  be  very  useful:    An  es.mated  one  million  Russian  motorists  have  dashboard  video  cameras  installed  in  their  cars.    

 Police  officers  carring  video  camera  units  and  using  Velcro  to  place  these  cameras  in  police  wagons,  helmet  cams,  ear  cams,  chest  cams  with  audio  capability,  GPS  locators,  taser  cams    

•  “Even  with  only  half  of  the  54  uniformed  patrol  officers  wearing  cameras  at  any  given  .me,  a  department  in  USA  had  an  88  %  decline  in  the  number  of  complaints  filed  against  officers,  compared  with  the  12  months  before  the  study”,  The  New  York  Times,  4th  of  July,  2013.  

       

The  world’s  leading  police  body  worn  video  camera  deployed  by  over  4000  agencies  in  16  countries.  

Page 23: CAIP 2015 Introduction on Visual Lifelogging I part

Benefits  and  poten.al  applica.ons  

ì  It  will  take  quite  some  .me  for  people  to  feel  comfortable  with    ‘always  connected’  devices  that  can  discreetly  take  photos  or  videos.  Will  the  benefits  outweigh  the  nega.ves?  

“Quan.fied  Self  &  life-­‐logging  Meets  Internet  of  Things  (IOT)”,  Dr.  Mazlan  Abbas,  MIMOS  Berhad  

23  

Page 24: CAIP 2015 Introduction on Visual Lifelogging I part

Life-­‐logging  and  Egocentric  Vision  

Page 25: CAIP 2015 Introduction on Visual Lifelogging I part

Egocentric  vision  “Taking  photos  can  impair  your  ability  to  remember”    

Next  4me  you're  at  a  museum  or  an  event,  think  before  you  snap  a  photo!  

“SenseCam  has  already  made  an  impact  in  memory  enhancement”    

 

ì  What  do  we  have?  ì  Self  centric,  automa.cally  captured  Images  of  our  life  ì  Objec3ve  photos  versus  Subjec3ve  (tradi.onal)  photos  ì  Most  similar  photos  to  our  memory  

25  

Page 26: CAIP 2015 Introduction on Visual Lifelogging I part

Egocentric  vision  “Taking  photos  can  impair  your  ability  to  remember”    

Next  4me  you're  at  a  museum  or  an  event,  think  before  you  snap  a  photo!  

“SenseCam  has  already  made  an  impact  in  memory  enhancement”    

ì  What  do  we  have?  ì  Self  centric,  automa.cally  captured  Images  of  our  life  ì  Objec3ve  photos  versus  Subjec3ve  (tradi.onal)  photos  ì  Most  similar  photos  to  our  memory  

ì  What  we  are  looking  for?  ì  Seman3c  informa.on  (life-­‐logs,  NOT  data)  

ì  A  Search  Engine  for  the  Self  is  required!  (life-­‐logging)  

26  

Page 27: CAIP 2015 Introduction on Visual Lifelogging I part

Egocentric  vision  

ì  What  we  want:  

 

Events  to  be  extracted  from  life-­‐logging  images  

27  

Page 28: CAIP 2015 Introduction on Visual Lifelogging I part

Egocentric  vision  

28  

ì  What  we  have:  

 

Page 29: CAIP 2015 Introduction on Visual Lifelogging I part

Wealth  of  life-­‐logging  data  ì  We  propose  an  energy-­‐based  approach  for  mo.on-­‐based  event  

segmenta.on  of  life-­‐logging  sequences  of  low  temporal  resolu.on  

ì  -­‐  The  segmenta.on  is  reached  integra.ng  different  kind  of  image  features  and  classifiers  into  a  graph-­‐cut  framework  to  assure  consistent  sequence  treatment.  

Complete  dataset  of  a  day  captured  with  SenseCam  (more  than  4,100  images  

IMEC,  University  of  Eindhoven,  March,  2015   29  

Choice  of  devise  depends  on:    1)  where    they  are  set:  a  hung  up  camera  has  the  advantage  that  is  considered  more  unobtrusive  for  the  user,  or    2)  their  temporal  resolu.on:  a  camera  with  a  low  fps  will  capture  less  mo.on  informa.on,  but  we  will  need  to  process  less  data.  We  chose  a  SenseCam  or  Narra.ve  -­‐  cameras  hung  on  the  neck  or  pinned  on  the  dress  that  capture  2-­‐4  fps.  

Or  the  hell  of  life-­‐logging  data  

Page 30: CAIP 2015 Introduction on Visual Lifelogging I part

Towards  lifestyle  characteriza.on  We  want  to  extract  life-­‐logging  informa3on  about:  

ì  Events    

ì  Ac.vi.es  

ì  Social  interac.ons,  etc..  

ì  Memorable  moments  

ì  Personal  habits  and  context  

ì  Lifestyle…  healthy  lifestyle…  

30  

Our  egocentric  vision  research:  

ì  Video  segmenta.on  for  events  extrac.on    

ì  Mo.on-­‐based  video  segmenta.on  towards  ac.vi.es  recogni.on  

ì  Human  tracking,  towards  social  interac.on  and  key-­‐frame  extrac.on  

ì  Ac.ve  learning  for  object  recogni.on    

ì  Object  discovery  for  lifestyle  characteriza.on,  etc,  etc.  

What?  Where?  When?  Who?  

Page 31: CAIP 2015 Introduction on Visual Lifelogging I part

Egocentric  vision:  an  overview  

31  Marc  Bolaños,  Mariella  Dimiccoli,  Pe3a  Radeva,  “Towards  Storytelling  from  Visual  Lifelogging:  An  Overview”,  arXiv:1507.06120,  2015,  (submiaed  to  a  Journal).  

Page 32: CAIP 2015 Introduction on Visual Lifelogging I part

Informa.veness  CNN  

Day  Lifelog  Informa.ve  Images  

Egocentric  informa.veness  through  deep  learning  

Page 33: CAIP 2015 Introduction on Visual Lifelogging I part

Temporal  video  segmenta.on  

33  

Life-­‐logging   Video  Segments  Extrac3on  Data  set:  thousands  of  

images  

Features  Extrac.on   Clustering   Daily  Video  

Summariza.on  

Event  extrac.on  Event  

characteriza.on  

RGB+HOG  CNN  

Ward  k-­‐Means  

Spectral  Clustering  

Estefania  Talavera,  Mariella  Dimiccoli,  Marc  Bolaños,  Maedeh  Aghaei  and  Pe3a  Radeva.  R-­‐clustering  for  Egocentric  Video  Segmenta3on,  IBPRIA’15,  San3ago  de  Compostela,  2015,  (extended  version  submiaed  to  a  journal).  

Page 34: CAIP 2015 Introduction on Visual Lifelogging I part

Mo.on-­‐based  video  segmenta.on  ì  Event  segmenta.on  in  LL  data  is  characterized  by  the  movement  of  the  

person  wearing  the  device.  

ì  Group  consecu.ve  frames  in  three  general  event  classes:  •  ”Sta.c”  (person  is  not  

moving)    •  ”In  Transit”  (person  is  moving  

or  running)    •  ”Moving  Camera”  (person  is  

in  the  same  area  performing  some  ac.on).  

34  M.  Bolaños,  M.  Garolera,  P.  Radeva,  “Video  segmenta3on  of  life-­‐logging  videos”,  AMDO’2014,  Springer-­‐Verlag,  2014,  p.1-­‐9.  

Page 35: CAIP 2015 Introduction on Visual Lifelogging I part

Visual  summary  through  keyframe  extrac.on  

M  Bolaños,  R  Mestre,  E  Talavera,  X  Giró-­‐i-­‐Nieto,  P  Radeva,  VISUAL  SUMMARY  OF  EGOCENTRIC  PHOTOSTREAMS  BY  REPRESENTATIVE  KEYFRAMES,  arXiv  preprint  arXiv:1505.01130,  2015  IEEE  Interna3onal  Conference  on  Mul3media  and  ExpoWorkshops    (ICMEW).    

35  

Results  

Joint  work  btw  UPC  &  UB.  

Page 36: CAIP 2015 Introduction on Visual Lifelogging I part

Seman.c  Summariza.on  

•  Images  should  be:  

1.   Not  repe33ve  

2.   Not-­‐usual  

•  And  answering  to:  

1.   Who?  

2.   What?  

3.   How?  

4.   When?  

5.   Where?  

In  order  to  construct  the  storytelling:  

Results:  

We  define  a  criterion  to  choose  the  keyframes  based  on  saliency,  objectness  and  faceness:  

Master  of  Aniol  lLidon,  supervised  by  Xavier  Giro-­‐i-­‐Nieto  a¡&  Pe3a  Radeva   Joint  work  btw  UPC  &  UB.  

Page 37: CAIP 2015 Introduction on Visual Lifelogging I part

Food  detec.on  in  egocentric  vision  How  can  we  automa.cally  detect  every  instance  of  a  dish  in  all  of  its  variants,  shapes  and  posi3ons  and  in  such  a  large  number  of  images?    The  main  problems  that  arise  are:  •  Complexity  and  variability  of  the  data.  •  Huge  amounts  of  data  to  analyse  (up  to  100.000  images  per  month).    Any  efficient  supervised  classifier  needs  a  huge  amount  of  training  set!    Using  human  .me  for  labelling  million  of  images  is  too  expensive.          

37  

We  need  an  automa3c  aid  for  labelling  huge  amount  of  images.  

Page 38: CAIP 2015 Introduction on Visual Lifelogging I part

Ac.ve  learning  for  nutri.on-­‐related  objects  annota.on  

38  Marc  Bolaños,  Maite  Garolera,  Pe.a  Radeva,  "Ac.ve  Labeling  Applica.on  Applied  to  Food-­‐Related  Object  Recogni.on",  5th  Workshop  on  Mul.media  for  Cooking  and  Ea.ng  Ac.vi.es  :  CEA2013,  ACM  Interna.onal  Conference  on  Mul.media  2013,  Barcelona  October,  2013.  

Page 39: CAIP 2015 Introduction on Visual Lifelogging I part

Ac.ve  Learning  

ì  Goal:  Minimize  human  effort  by  guiding  the  process  of  crea.ng  the  training  set.  

39  

Effect  of  ac.ve  learning:  improvement  of  learning  performance  (ley),  improvement  of  training  .me  (right).    

Page 40: CAIP 2015 Introduction on Visual Lifelogging I part

Object  discovery  for  context  charaterisa.on  ì  What  is  the  context  of  a  person  day-­‐by-­‐day?  

Do  these  sets  belong  to  the  same  person?  

Person  1  

Person  1  

Person  2  

M.  Bolaños,  P.  Radeva,  “Object  Discovery  for  Egocentric  Videos  Based  on  Convolu.onal  Neural  Network  Features”,  Workshop  on  Story  telling,  ECCV,  2014.      Marc  Bolaños,  Maite  Garolera  and  Pe.a  Radeva.  Object  Discovery  using  CNN  Features  in  Egocentric  Videos,  IBPRIA’15,  San.ago  de  Compostela,  2015.  (submibed  to  a  journal)      

40  

Page 41: CAIP 2015 Introduction on Visual Lifelogging I part

Social  analysis  ì  Who  was  with  us?  For  how  long?  

ì  Humans  are  important  

ì  Video  segmenta3on  based  on  presence  of  people  ì  Each  segment  is  an  event  related  to  specific  person  

ì  Intui.ve  solu.on  is  to  track  visible  people  

41  

Page 42: CAIP 2015 Introduction on Visual Lifelogging I part

Social  analysis  

ì  A  tracking  method  independent  from  background  changes  and  temporal  resolu3on  could  fit  this  data  

42  M.  Aghaei,  P.  Radeva,  “Bag-­‐of-­‐Tracklets  for  Person  Tracking  in  Life-­‐Logging  Data”,  CCIA’2014.  Extended  version  submibed  to  a  journal.  

Page 43: CAIP 2015 Introduction on Visual Lifelogging I part

Analysis  of  human  interac.on  

 Maedeh  Aghaei,  Mariella  Dimiccoli,  Pe3a  Radeva,  “Towards  Social  Interac3on  Detec3on  in  Egocentric  Photo-­‐streams”,  (submiaed)  43  

!

F-­‐forma.ons  extrac.on  from  egocentric  vision  

Page 44: CAIP 2015 Introduction on Visual Lifelogging I part

44  

Page 45: CAIP 2015 Introduction on Visual Lifelogging I part

Life-­‐logging  analysis  for  MCI  treatment  

45  

Goal:  to  develop  tools  for  memory  reforcing  of  MCI  and  Alzheimer  people.  •  To  develop  a  life-­‐logging  system  recording  specific  autobiographical  episodes  for  s3mula3ng  posteriorly  episodic  memory  func3on  

known  to  be  deficient  in  MCI.    

•  To  explore  the  associa.on  btw  biomarkers  changes  in  cogni3ve,  func3onal  and  emo3onal  outcomes.  

•  To  learn  more  about  the  underlying  biological  mechanisms  for  how  effec3ve  behavioural  interven3ons  improve  cogni3ve  and  func3onal  outcomes.    

Inspired  in  the  case  of  use  of  SenseCam  for  a  pa.ent  with  herpes  encephali.s.  

But  what  are  the  memorable  moments?  

Page 46: CAIP 2015 Introduction on Visual Lifelogging I part

Life-­‐logging  for  wellfare  

 •   how  to  extract  seman3c  units  related  to  the  lifestyle  and  their  context  rela.on,  •     • how  to  segment  life-­‐log  data  into  meaningful  events,    

• what  are  the  seman.c  units  that  characterize  the  lifestyle  of  individuals,    

• what  is  their  rela3on  and  how  the  context  affects  them,    

• how  to  extract  and  characterize  lifestyles  paberns,  

•   what  is  the  healtstyle,  etc.    

To  derive  lifestyle  paberns  from  visual  life-­‐logs  and  to  conduct  a  study  on  the  feasibility  of  automa3cally  genera3on  of  lifestyle  paaerns  and  interpreta3ons  to  be  used  in  the  future  to  improve  lifestyle  of  individuals.    

46  

Page 47: CAIP 2015 Introduction on Visual Lifelogging I part

Towards  lifestyle  characteriza.on  Towards  visualizing  summarized  lifestyle  data  to  ease  the  management  of  the  user’s  healthy  habits  (sedentary  lifestyles,  nutri.onal  ac.vity,  etc.).  

 

47  

Life-­‐logging  can  help  us  accomplish  our  goals:  taking  photos  of  our  everyday  life  and  being  able  to  analyse  what  we  eat,  star.ng  by  the  dish  recogni.on.    

 Aim:  to  develop  tools  based  on  lifelogging  to  show  the  rela.on  between  nutri.on  and  cogni.on.  European  proposal  submibed  to  the  call  in  April  2015.    

Page 48: CAIP 2015 Introduction on Visual Lifelogging I part

Conclusions  ì  Life-­‐logging  is  a  very  recent  trend  growing  very  fast  and  with  a  huge  poten.al.  

ì  Technology  for  data  acquisi.on  and  storage  is  ready.    

ì  Life-­‐loggers  are  wai3ng  for  tools  of  egocentric  vision  to  process  the  huge  amount  of  data.  

ì  We  developed  several  algorithms  for:  ì  Video  segmenta.on  

ì  Event  detec.on  

ì  Object  and  scene  recogni.on    

ì  Key  frame  extrac.on  ì  Video  summarization  and  browsing  

ì  Social  interac.on  

ì  Storytelling  

ì  Huge  poten3al  for  numerous  applica3ons  to  health  &  wellfare  that  will  strongly  influence  the  future  of  the  technology    

48