Picture Fusion for Context Enhancement and Video Surrealism

Slide1 l.jpg
1 / 35
0
0
1239 days ago, 328 views
PowerPoint PPT Presentation
. . Dim Bldgs. Reflections on bldgs. Obscure shapes. .

Presentation Transcript

Slide 1

Picture Fusion for Context Enhancement and Video Surrealism Ramesh Raskar Mitsubishi Electric Research Labs, (MERL) Adrian Ilie UNC Chapel Hill Jingyi Yu MIT

Slide 3

Dark Bldgs Reflections on bldgs Unknown shapes

Slide 4

'Sufficiently bright' Bldgs Reflections in bldgs windows Tree, Street shapes

Slide 5

Night Image Background is caught from day-time scene utilizing the same settled camera Context Enhanced Image Day Image

Slide 6

Mask is consequently figured from scene differentiate

Slide 7

But, Simple Pixel Blending Creates Ugly Artifacts

Slide 8

Pixel Blending

Slide 9

Pixel Blending Our Method : Integration of mixed Gradients

Slide 10

Outline Context Enhancement Gradient-based Fusion Video Enhancement Surrealism

Slide 11

Gradient field Nighttime picture x Y I 1 G 1 G 1 Mixed inclination field x Y G Importance picture W I 2 x Y G 2 G 2 Final outcome Daytime picture Gradient field

Slide 12

Reconstruction from Gradient Field Problem: limit blunder | Ñ I' – G| Estimate I' so that G = Ñ I' Poisson condition Ñ 2 I' = div G Full multigrid solver G X I' G Y

Slide 13

Why Gradient-based Approach Comparison of force qualities are vital Maintain angles to catch nearby varieties Directly understand for craved slopes Maintain unobtrusive points of interest Mix divergent pictures No requirement for exact division

Slide 14

Comparison Average Subtle points of interest are lost Pixel-wise mixing Sharp moves

Slide 15

Issues Boundary conditions Color movements

Slide 16

Boundary Conditions Assumed Neumann condition at outskirts, Ñ I' · N = 0 , Enforced by haloing picture with blacks

Slide 17

Color Shift Mixing disparate pictures Goal: last picture appearance matches input pictures at relating pixels I last (x,y) = c 1 I poisson (x,y) + c 2 Solve  W i (x,y) I unique (x,y) = c 1 I poisson (x,y) + c 2 Each shading channel remade independently

Slide 19

Outline Context Enhancement Gradient-based Fusion Video Enhancement Surrealism

Slide 22

Overview of Process Original evening activity camera 320x240 video Day time picture: By averaging 5 seconds of day video Input Output Enhanced video Note: off-ramp, path dividers, structures not unmistakable in unique night video, but rather plainly observed here. Cover outline (for edge above): Encodes pixel with force change

Slide 23

Algorithm Frame N Gradient field Mixed inclination field TimeAveraged significance veil Processed twofold cover Final outcome Gradient field Frame N-1 Daytime picture

Slide 24

Outline Context Enhancement Gradient-based Fusion Video Enhancement Related Work Surrealism

Slide 25

Related Work Spatio-worldly Composition Duchamp ( Nude dropping a staircase ) Freeman 2002 Fels 1999, Klein 2002, Cohen 2003 Gradient-based Techniques Multi-unearthly: Socolinsky 1999 Shadow expulsion: Weiss 2001 High element extend: Fattal 2002 Image altering: Perez 2003 Some at Siggraph'04

Slide 26

Surrealism Rene Magritte, 'Domain of the Light'

Slide 27

Outline Context Enhancement Gradient-based Fusion Video Enhancement Surrealism

Slide 28

Time-slip by Mosaics Maggrite Stripes time

Slide 29

Time Lapse Mosaic

Slide 30

Time Lapse Mosaic

Slide 31

t

Slide 32

Sunrise at Night

Slide 33

BiSolar System

Slide 34

Discussion User Experience More viable in passing on scene setting "Marvelous" appearance Nonrealistic : False conditions Applications Tools for specialists Surveillance Amusement stop rides Performance ~1 sec/outline for 320x240 ~ 3 min for 4Mpixel picture

Slide 35

Image Fusion for Context Enhancement Nonrealistic however fathomable setting Fusion utilizing numerous pictures Enhancing night pictures with day bgrnd Gradient-based combination Video surrealism instruments t

SPONSORS