emer emergenc ence o e of o f obj bjec ect s seg egmen
play

Emer Emergenc ence o e of O f Obj bjec ect S Seg egmen - PowerPoint PPT Presentation

Emer Emergenc ence o e of O f Obj bjec ect S Seg egmen mentatio ion n in P in Per ertur urbed Gener bed Generativ ive M e Models dels Adam Bielski, Paolo Favaro University of Bern Ob Object se segm gmentation Segmentation


  1. Emer Emergenc ence o e of O f Obj bjec ect S Seg egmen mentatio ion n in P in Per ertur urbed Gener bed Generativ ive M e Models dels Adam Bielski, Paolo Favaro University of Bern

  2. Ob Object se segm gmentation Segmentation model 2

  3. Ob Object se segm gmentation Annotated data Segmentation model 3

  4. Ob Object se segm gmentation Annotated data Segmentation model • Human annotated • Expensive • Expertise knowledge 4

  5. Un Unsuper ervis vised ed objec ject t seg egmen entatio tion Annotated data Segmentation model • Human annotated • Expensive • Expertise knowledge Can we do it without labels? 5

  6. Un Unsuper ervis vised ed ob object se segm gmentation E G Mask Image 6

  7. Le Learn rning to to ge generate la layer ered ed sc scenes wi with GA GAN Background Foreground E G Mask Image 𝑨~𝒪(0, 𝐽) 7

  8. Le Learn rning to to ge generate la layer ered ed sc scenes wi with GA GAN Background Image composition Foreground + ⊙ ⊙ = E G Background 1 - Mask Foreground Mask Mask Composite image Image 𝑨~𝒪(0, 𝐽) 8

  9. Le Learn rning to to ge generate la layer ered ed sc scenes wi with GA GAN Background Image composition Foreground + ⊙ ⊙ = E G Background 1 - Mask Foreground Mask Mask Composite image Image 𝑨~𝒪(0, 𝐽) Real / D Fake Unlabeled real images 9

  10. De Degenerate s solu lutio ions Background Foreground Mask Composite image 10

  11. De Degenerate s solu lutio ions Background Foreground Mask Composite image 11

  12. De Degenerate s solu lutio ions Background Foreground Mask Composite image 12

  13. De Degenerate s solu lutio ions Background Foreground Mask Composite image 13

  14. Wha What is is a c a correct ect par partit itio ion? n? 14

  15. Wha What is is a c a correct ect par partit itio ion? n? 15

  16. Ra Rando ndom m shi hifts 16

  17. Ra Rando ndom m shi hifts 17

  18. Ra Rando ndom m shi hifts 18

  19. Ra Rando ndom m shi hifts 19

  20. Invalid partition ⟹ invalid scene after a small shift Valid partition ⟹ valid scene after a small shift Ra Rando ndom m shi hifts 20

  21. Av Avoiding degenerate solutions Background Foreground Mask Composite image Composite image w/shift 21

  22. Av Avoiding degenerate solutions Background Foreground Mask Composite image Composite image w/shift 22

  23. Av Avoiding degenerate solutions Background Foreground Mask Composite image Composite image w/shift 23

  24. Av Avoiding degenerate solutions Background Foreground Mask Composite image Composite image w/shift 24

  25. Av Avoiding degenerate solutions Background Foreground Mask Composite image Composite image w/shift Add a loss term to avoid empty masks 25

  26. Av Avoiding degenerate solutions Background Foreground Mask Composite image Composite image w/shift Add a loss term to avoid empty masks 26

  27. Learn Le rning wi with per pertur urba bations ns Background Image composition Foreground + ⊙ ⊙ = E G Background 1 - Mask Foreground Mask Mask Composite image Image 𝑨~𝒪(0, 𝐽) Minimum mask loss 27

  28. Learn Le rning wi with per pertur urba bations ns Background Image composition Foreground + ⊙ ⊙ = E G Background 1 – Mask Foreground Mask Mask Composite image Image shifted shifted shifted (fake) 𝑨~𝒪(0, 𝐽) Random shift 𝑞~𝒱 Real Minimum mask loss / D Fake Real image 28

  29. Ge Generatio tion results lts LSUN Car LSUN Bird LSUN Chair LSUN Horse Backg�o�nd Backg�o�nd Fo�eg�o�nd Fo�eg�o�nd Ma�k Ma�k Fo�eg�o�nd Fo�eg�o�nd �/�ma�k �/�ma�k Com�o�i�e Com�o�i�e image image 29

  30. Le Learn rning to to se segm gment Background Image composition Foreground + ⊙ ⊙ = G Background 1 - Mask Mask Mask Foreground Composite image 30

  31. Learn Le rning to to se segm gment Frozen Background Image composition Foreground + ⊙ ⊙ = G Background 1 - Mask Mask Mask Foreground Composite image 31

  32. Le Learn rning to to se segm gment Frozen Background Image composition Foreground + ⊙ ⊙ = E G Background 1 - Mask Mask Mask Foreground Composite image Image 𝑀1 𝑀𝑝𝑡𝑡 32

  33. Le Learn rning to to se segm gment Frozen Background Image composition Foreground + ⊙ ⊙ = E G Background 1 - Mask Mask Mask Foreground Composite image Image 𝑀1 𝑀𝑝𝑡𝑡 33

  34. Se Segme mentation on r results Real image Real image O�� Our results me�hod Approx. Ma�k R-CNN ground truth Real image Real image Our Our results method Ground Ground truth truth 34

  35. Po Poster #60 E ast Exhibition Hall B + C 35

Recommend


More recommend