Context for single object classes<br />
Who needs context anyway?We can recognize objects even out of context<br />Banksy<br />
Why is context important?<br /><ul><li> Changes the interpretation of an object (or its function)
 Context defines what an unexpected event is </li></li></ul><li>Look-Alikes by Joan Steiner<br />Even in high resolution, ...
The importance of context<br />Cognitive psychology<br />Palmer 1975 <br />Biederman 1981<br />…<br />Computer vision<br /...
What is the context for a single object category?<br />
The influence of an object extends beyond its physical boundaries<br />
Global and local representations<br />building<br />Urban street scene<br />car<br />sidewalk<br />
Global and local representations<br />building<br />Urban street scene<br />car<br />sidewalk<br />Image index: Summary st...
Global scene representations<br />Spatially organized textures<br />Bag of words<br />M. Gorkani, R. Picard, ICPR 1994<br ...
S<br />g<br />An integrated model of Scenes, Objects, and Parts<br />Scene<br />Ncar<br />P(Ncar | S = street)<br />N<br /...
S<br />g<br />Context driven object detection<br />Scene<br />Zcar<br />Ncar<br />P(Ncar | S = street)<br />N<br />1<br />...
car<br />Fi<br />dcari<br />xcari<br />An integrated model of Scenes, Objects, and Parts<br />We train a multiview car det...
S<br />car<br />Fi<br />g<br />dcari<br />xcari<br />An integrated model of Scenes, Objects, and Parts<br />Scene<br />Zca...
A car out of context …<br />
~6cm<br />We are wired for 3D<br />
We can not shut down 3D perception<br />(c) 2006 Walt Anthony<br />
Scenes rule over objects<br />3D percept is driven by the scene, which imposes its ruling to the objects<br />
3D from pixel values<br />D. Hoiem, A.A. Efros, and M. Hebert, "Automatic Photo Pop-up”. SIGGRAPH 2005.<br />A. Saxena, M....
Surface Estimation<br />Object<br />Surface?<br />Support?<br />Image<br />Support<br />Vertical<br />Sky<br />V-Center<br...
Object Support<br />Slide by Derek Hoiem<br />
Slide by James Coughlan<br />
Slide by James Coughlan<br />
3d Scene Context<br />Image<br />World<br />Hoiem, Efros, Hebert ICCV 2005<br />
meters<br />meters<br />3D scene context<br />Ped<br />Ped<br />Car<br />Hoiem, Efros, Hebert ICCV 2005<br />
Qualitative Results<br />Car: TP / FP  Ped: TP / FP<br />Initial: 2 TP / 3 FP<br />Final: 7 TP / 4 FP<br />Local Detector ...
3D City Modeling using Cognitive Loops<br />N. Cornelis, B. Leibe, K. Cornelis, L. Van Gool.CVPR'06<br />
Single view metrology<br />Criminisi, et al. 1999<br /> Need to recover:<br /><ul><li> Ground plane
Upcoming SlideShare
Loading in …5
×

Iccv2009 recognition and learning object categories p1 c02 - detecting single objects in context

464 views

Published on

Published in: Education
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
464
On SlideShare
0
From Embeds
0
Number of Embeds
2
Actions
Shares
0
Downloads
10
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

Iccv2009 recognition and learning object categories p1 c02 - detecting single objects in context

  1. 1. Context for single object classes<br />
  2. 2. Who needs context anyway?We can recognize objects even out of context<br />Banksy<br />
  3. 3. Why is context important?<br /><ul><li> Changes the interpretation of an object (or its function)
  4. 4. Context defines what an unexpected event is </li></li></ul><li>Look-Alikes by Joan Steiner<br />Even in high resolution, we can not shut down contextual processing and it is hard to recognize the true identities of the elements that compose this scene.<br />
  5. 5.
  6. 6.
  7. 7.
  8. 8.
  9. 9. The importance of context<br />Cognitive psychology<br />Palmer 1975 <br />Biederman 1981<br />…<br />Computer vision<br />Noton and Stark (1971)<br />Hanson and Riseman (1978)<br />Barrow & Tenenbaum (1978) <br />Ohta, kanade, Skai (1978)<br />Haralick (1983)<br />Strat and Fischler (1991)<br />Bobick and Pinhanez (1995)<br />Campbell et al (1997)<br />
  10. 10. What is the context for a single object category?<br />
  11. 11. The influence of an object extends beyond its physical boundaries<br />
  12. 12. Global and local representations<br />building<br />Urban street scene<br />car<br />sidewalk<br />
  13. 13. Global and local representations<br />building<br />Urban street scene<br />car<br />sidewalk<br />Image index: Summary statistics, <br />configuration of textures<br />Urban street scene<br />histogram<br />features<br />
  14. 14. Global scene representations<br />Spatially organized textures<br />Bag of words<br />M. Gorkani, R. Picard, ICPR 1994<br />A. Oliva, A. Torralba, IJCV 2001<br />Sivic et. al., ICCV 2005<br />Fei-Fei and Perona, CVPR 2005<br />Non localized textons<br />…<br />Walker, Malik. Vision Research 2004 <br />…<br />S. Lazebnik, et al, CVPR 2006<br />Spatial structure is important in order to provide context for object localization<br />
  15. 15. S<br />g<br />An integrated model of Scenes, Objects, and Parts<br />Scene<br />Ncar<br />P(Ncar | S = street)<br />N<br />1<br />5<br />0<br />P(Ncar | S = park)<br />Scene<br />gist<br />features<br />N<br />1<br />5<br />0<br />
  16. 16. S<br />g<br />Context driven object detection<br />Scene<br />Zcar<br />Ncar<br />P(Ncar | S = street)<br />N<br />1<br />5<br />0<br />Scene<br />gist<br />features<br />
  17. 17. car<br />Fi<br />dcari<br />xcari<br />An integrated model of Scenes, Objects, and Parts<br />We train a multiview car detector. <br />p(d | F=1) = N(d | m1, s1)<br />p(d | F=0) = N(d | m0, s0)<br />N=4<br />
  18. 18. S<br />car<br />Fi<br />g<br />dcari<br />xcari<br />An integrated model of Scenes, Objects, and Parts<br />Scene<br />Zcar<br />Ncar<br />Scene<br />gist<br />features<br />M=4<br />P(F,S | x,d,g) a p(F | S)p(S | g) p(xi | g) PN(xi; mb, sb2) PN(di; mtp, stp2) PN(di; mtn, stn2)<br />i:Fi=0<br />i:Fi=0<br />i:Fi=1<br />
  19. 19.
  20. 20.
  21. 21. A car out of context …<br />
  22. 22. ~6cm<br />We are wired for 3D<br />
  23. 23. We can not shut down 3D perception<br />(c) 2006 Walt Anthony<br />
  24. 24. Scenes rule over objects<br />3D percept is driven by the scene, which imposes its ruling to the objects<br />
  25. 25. 3D from pixel values<br />D. Hoiem, A.A. Efros, and M. Hebert, "Automatic Photo Pop-up”. SIGGRAPH 2005.<br />A. Saxena, M. Sun, A. Y. Ng. "Learning 3-D Scene Structure from a Single Still Image"<br />In ICCV workshop on 3D Representation for Recognition (3dRR-07), 2007.<br />
  26. 26. Surface Estimation<br />Object<br />Surface?<br />Support?<br />Image<br />Support<br />Vertical<br />Sky<br />V-Center<br />V-Right<br />V-Porous<br />V-Solid<br />V-Left<br />[Hoiem, Efros, Hebert ICCV 2005]<br />Slide by Derek Hoiem<br />
  27. 27. Object Support<br />Slide by Derek Hoiem<br />
  28. 28. Slide by James Coughlan<br />
  29. 29. Slide by James Coughlan<br />
  30. 30. 3d Scene Context<br />Image<br />World<br />Hoiem, Efros, Hebert ICCV 2005<br />
  31. 31. meters<br />meters<br />3D scene context<br />Ped<br />Ped<br />Car<br />Hoiem, Efros, Hebert ICCV 2005<br />
  32. 32. Qualitative Results<br />Car: TP / FP Ped: TP / FP<br />Initial: 2 TP / 3 FP<br />Final: 7 TP / 4 FP<br />Local Detector from [Murphy-Torralba-Freeman 2003]<br />Slide by Derek Hoiem<br />
  33. 33. 3D City Modeling using Cognitive Loops<br />N. Cornelis, B. Leibe, K. Cornelis, L. Van Gool.CVPR'06<br />
  34. 34. Single view metrology<br />Criminisi, et al. 1999<br /> Need to recover:<br /><ul><li> Ground plane
  35. 35. Reference height
  36. 36. Horizon line
  37. 37. Where objects contact the </li></ul> ground<br />

×