本章节介绍了计算机视觉分割的基本概念,从两个方面来,分割和聚合,介绍了分割和聚合的具体知识,怎么来实施这两步骤,另外关于分割部分,介绍了知觉组织的格式塔理论,关于这一理论的由来以及如何去利用这个理论。

注脚

展开查看详情

1. Lecture  12:  Clustering  and   Segmenta5on   Professor  Fei-­‐Fei  Li   Stanford  Vision  Lab   Fei-Fei Li! Lecture 12 - 1  ! 28-­‐Oct-­‐14  

2. What  we  will  learn  today   •  Introduc5on  to  segmenta5on  and  clustering   •  Gestalt  theory  for  perceptual  grouping   •  Agglomera5ve  clustering   Reading:  [FP]  Chapters:  14.2,  14.4     Fei-Fei Li! Lecture 12 - 2  ! 28-­‐Oct-­‐14  

3.Fei-Fei Li! Lecture 12 - 3  ! 28-­‐Oct-­‐14  

4. Image  Segmenta5on   •  Goal:  iden5fy  groups  of  pixels  that  go  together   Slide credit: Steve Seitz, Kristen Grauman Fei-Fei Li! Lecture 12 - 4  ! 28-­‐Oct-­‐14  

5. The  Goals  of  Segmenta5on   •  Separate  image  into  coherent  “objects”   Image   Human  segmenta5on   Slide credit: Svetlana Lazebnik Fei-Fei Li! Lecture 12 - 5  ! 28-­‐Oct-­‐14  

6. The  Goals  of  Segmenta5on   •  Separate  image  into  coherent  “objects”   •  Group  together  similar-­‐looking  pixels  for   efficiency  of  further  processing   Slide credit: Svetlana Lazebnik “superpixels” X.  Ren  and  J.  Malik.  Learning  a  classifica5on  model  for  segmenta5on.  ICCV  2003.   Fei-Fei Li! Lecture 12 - 6  ! 28-­‐Oct-­‐14  

7. Segmenta5on  for  feature  support   50x50 Patch 50x50 Patch Slide:  Derek  Hoiem   Fei-Fei Li! Lecture 12 - 7  ! 28-­‐Oct-­‐14  

8. Segmenta5on  for  efficiency   [Felzenszwalb and Huttenlocher 2004] [Hoiem et al. 2005, Mori 2005] [Shi and Malik 2001] Slide:  Derek  Hoiem   Fei-Fei Li! Lecture 12 - 8  ! 28-­‐Oct-­‐14  

9. Segmenta5on  as  a  result   Rother et al. 2004 Fei-Fei Li! Lecture 12 - 9  ! 28-­‐Oct-­‐14  

10. Types  of  segmenta5ons   Oversegmentation Undersegmentation Multiple Segmentations Fei-Fei Li! Lecture 12 - 10   ! 28-­‐Oct-­‐14  

11. One  way  to  think  about   “segmenta5on”  is  Clustering        Clustering:  group  together  similar  points  and   represent  them  with  a  single  token      Key  Challenges:    1)  What  makes  two  points/images/patches  similar?    2)  How  do  we  compute  an  overall  grouping  from   pairwise  similari5es?     Slide:  Derek  Hoiem   Fei-Fei Li! Lecture 12 - 11   ! 28-­‐Oct-­‐14  

12. Why  do  we  cluster?   •  Summarizing  data   –  Look  at  large  amounts  of  data   –  Patch-­‐based  compression  or  denoising   –  Represent  a  large  con5nuous  vector  with  the  cluster  number     •  Coun2ng   –  Histograms  of  texture,  color,  SIFT  vectors   •  Segmenta2on   –  Separate  the  image  into  different  regions   •  Predic2on   –  Images  in  the  same  cluster  may  have  the  same  labels   Slide:  Derek  Hoiem   Fei-Fei Li! Lecture 12 - 12   ! 28-­‐Oct-­‐14  

13. How  do  we  cluster?   •  Agglomera5ve  clustering   –  Start  with  each  point  as  its  own  cluster  and   itera5vely  merge  the  closest  clusters   •  K-­‐means  (next  lecture)   –  Itera5vely  re-­‐assign  points  to  the  nearest  cluster   center   •  Mean-­‐shig  clustering  (next  lecture)   –  Es5mate  modes  of  pdf   •  Spectral  clustering  (CS231a,  winter  quarter)   –  Split  the  nodes  in  a  graph  based  on  assigned  links   with  similarity  weights   Fei-Fei Li! Lecture 12 - 13   ! 28-­‐Oct-­‐14  

14. General  ideas   •  Tokens   –  whatever  we  need  to  group  (pixels,  points,   surface  elements,  etc.,  etc.)   •  Bokom  up  clustering   –  tokens  belong  together  because  they  are  locally   coherent   •  Top  down  clustering   –  tokens  belong  together  because  they  lie  on  the   same  visual  en5ty  (object,  scene…)    >  These  two  are  not  mutually  exclusive   Fei-Fei Li! Lecture 12 - 14   ! 28-­‐Oct-­‐14  

15. Examples  of  Grouping  in  Vision   Grouping  video  frames  into  shots   Determining  image  regions   What things should Figure-­‐ground   be grouped? What cues indicate groups? Slide credit: Kristen Grauman Object-­‐level  grouping   Fei-Fei Li! Lecture 12 - 15   ! 28-­‐Oct-­‐14  

16. Similarity   Slide credit: Kristen Grauman Fei-Fei Li! Lecture 12 - 16   ! 28-­‐Oct-­‐14  

17. Symmetry   Slide credit: Kristen Grauman Fei-Fei Li! Lecture 12 - 17   ! 28-­‐Oct-­‐14  

18. Common  Fate   Image  credit:  Arthus-­‐Bertrand  (via  F.  Durand)   Slide credit: Kristen Grauman Fei-Fei Li! Lecture 12 - 18   ! 28-­‐Oct-­‐14  

19. Proximity   Slide credit: Kristen Grauman Fei-Fei Li! Lecture 12 - 19   ! 28-­‐Oct-­‐14  

20. Muller-­‐Lyer  Illusion   •  What  makes  the  bokom  line  look  longer  than   the  top  line?   Fei-Fei Li! Lecture 12 - 20   ! 28-­‐Oct-­‐14  

21. What  we  will  learn  today   •  Introduc5on  to  segmenta5on  and  clustering   •  Gestalt  theory  for  perceptual  grouping   •  Agglomera5ve  clustering   Fei-Fei Li! Lecture 12 - 21   ! 28-­‐Oct-­‐14  

22. The  Gestalt  School   •  Grouping  is  key  to  visual  percep5on   •  Elements  in  a  collec5on  can  have  proper5es  that   result  from  rela5onships     –  “The  whole  is  greater  than  the  sum  of  its  parts”   Illusory/subjec5ve     Occlusion   contours   Slide credit: Svetlana Lazebnik Familiar  configura5on   hkp://en.wikipedia.org/wiki/Gestalt_psychology   Fei-Fei Li! Lecture 12 - 22   ! 28-­‐Oct-­‐14  

23. Gestalt  Theory   •  Gestalt:  whole  or  group   –  Whole  is  greater  than  sum  of  its  parts   –  Rela5onships  among  parts  can  yield  new  proper5es/features   •  Psychologists  iden5fied  series  of  factors  that  predispose  set   of  elements  to  be  grouped  (by  human  visual  system)   “I stand at the window and see a house, trees, sky. Theoretically I might say there were 327 brightnesses and nuances of colour. Do I have "327"? No. I have sky, house, and trees.” Max Wertheimer (1880-1943) Untersuchungen zur Lehre von der Gestalt, Psychologische Forschung, Vol. 4, pp. 301-350, 1923 http://psy.ed.asu.edu/~classics/Wertheimer/Forms/forms.htm Fei-Fei Li! Lecture 12 - 23   ! 28-­‐Oct-­‐14  

24. Gestalt  Factors   Image source: Forsyth & Ponce •  These  factors  make  intui5ve  sense,  but  are  very  difficult  to  translate  into  algorithms.   Fei-Fei Li! Lecture 12 - 24   ! 28-­‐Oct-­‐14  

25. Con5nuity  through  Occlusion  Cues   Fei-Fei Li! Lecture 12 - 25   ! 28-­‐Oct-­‐14  

26. Con5nuity  through  Occlusion  Cues   Con5nuity,  explana5on  by  occlusion   Fei-Fei Li! Lecture 12 - 26   ! 28-­‐Oct-­‐14  

27. Con5nuity  through  Occlusion  Cues   Image source: Forsyth & Ponce Fei-Fei Li! Lecture 12 - 27   ! 28-­‐Oct-­‐14  

28. Con5nuity  through  Occlusion  Cues   Image source: Forsyth & Ponce Fei-Fei Li! Lecture 12 - 28   ! 28-­‐Oct-­‐14  

29. Figure-­‐Ground  Discrimina5on     Fei-Fei Li! Lecture 12 - 29   ! 28-­‐Oct-­‐14