SLIDE 1
Mosaics of Scenes with Moving Objects - CVPR98
Mosaics of Scenes with Moving Objects - CVPR98
Motivation
- Panoramic imagery
- Large high resolution images
Mosaics of Scenes with Moving Objects James Davis Computer - - PowerPoint PPT Presentation
Mosaics of Scenes with Moving Objects James Davis Computer - - PowerPoint PPT Presentation
Mosaics of Scenes with Moving Objects James Davis Computer Science Department Stanford University Mosaics of Scenes with Moving Objects - CVPR98 Motivation Panoramic imagery Large high resolution images Mosaics of Scenes with
SLIDE 2
SLIDE 3
Mosaics of Scenes with Moving Objects - CVPR98
Overview
Registration algorithms do not account for moving objects. Use phase correlation and estimate correct projective geometry. Mosaics with many images are ruined by accumulated registration errors. Find many local registrations and solve a linear system to obtain global registration. The final mosaic is blurry in regions of motion. Segment the mosaic into disjoint regions and fill each from a single source image.
SLIDE 4
Mosaics of Scenes with Moving Objects - CVPR98
Related Work
- Pairwise Registration
- ✁
- ✱
- Global Registration
- ✝
- ✝
- Moving Objects
- ✝
SLIDE 5
Mosaics of Scenes with Moving Objects - CVPR98
Pairwise Registration
- Fixed center of projection
- Parameter estimation
- Robust registration with moving objects
SLIDE 6
Mosaics of Scenes with Moving Objects - CVPR98
Phase correlation recovers translation
I2(x, y) = I1(x-x0, y-y0) δ(x-x0, y-y0) = F-1 F*[I2]F[I1] |F*[I2]F[I1]| x0
SLIDE 7
Mosaics of Scenes with Moving Objects - CVPR98
Phase correlation is not biased by moving objects
(a) (b) (c) (d)
SLIDE 8
Mosaics of Scenes with Moving Objects - CVPR98
Mellin transform recovers rotation and translation
- Extends phase correlation
- Polar transform converts rotation into
translation
- Assumes orthogonal projection
SLIDE 9
Mosaics of Scenes with Moving Objects - CVPR98
Finding the projection matrix
- Possess 2D parameters (x0, y0, θ0)
- Desire 3D Euler rotation angles (α, β, θ)
- Assume small angular rotation
- A = C-1RC
I2 f1 f2
α f x x
I1
) , arctan 2 , arctan 2 ( ) , , (
2 2
θ θ β α + = y x f f x
f x0 f x0
A : Image plane projection matrix C : Intrinsic camera matrix R : 3D rotation matrix
SLIDE 10
Mosaics of Scenes with Moving Objects - CVPR98
Without global registration errors accumulate
SLIDE 11
Mosaics of Scenes with Moving Objects - CVPR98
Relating pairwise and global registration
- Aij Pj = Pi
- Given all Aij , find all Pk
P1 A21 P2
Reference image I m a g e 1 Image 2
Aij : Pairwise projection of image i onto image j Pk : Projection of image k onto global reference plane
SLIDE 12
Mosaics of Scenes with Moving Objects - CVPR98
Globally registered mosaic
SLIDE 13
Mosaics of Scenes with Moving Objects - CVPR98
Compositing
- Blending produces blurring
- How do we avoid this?
SLIDE 14
Mosaics of Scenes with Moving Objects - CVPR98
Segment the mosaic
- A single source image per region
- Avoiding artifacts along boundaries?
SLIDE 15
Mosaics of Scenes with Moving Objects - CVPR98
Finding boundaries
- Avoid contradictory information
- Relative difference image
- Minimum difference path
SLIDE 16
Mosaics of Scenes with Moving Objects - CVPR98
Mosaic without blurring
SLIDE 17
Mosaics of Scenes with Moving Objects - CVPR98
Global registration comparison
SLIDE 18
Mosaics of Scenes with Moving Objects - CVPR98
Compositing comparison
SLIDE 19
Mosaics of Scenes with Moving Objects - CVPR98
Summary
- Contributions
SLIDE 20
Mosaics of Scenes with Moving Objects - CVPR98
Discussion
- Quantify small angle approximation
- Matrix elements do not have uniform scales
- Avoid segmenting into tiny regions
- Moving objects must appear in one image