9
Deliverable 2.2: Demonstrator on Multiscale Motion Estimation Florian Becker, Jing Yuan, Christoph Schnörr CVGPR group, University of Mannheim

Deliverable 2.2: Demonstrator on Multiscale Motion Estimation Florian Becker, Jing Yuan, Christoph Schnörr CVGPR group, University of Mannheim

Embed Size (px)

Citation preview

Page 1: Deliverable 2.2: Demonstrator on Multiscale Motion Estimation Florian Becker, Jing Yuan, Christoph Schnörr CVGPR group, University of Mannheim

Deliverable 2.2:

Demonstrator on Multiscale Motion Estimation

Florian Becker, Jing Yuan, Christoph Schnörr

CVGPR group, University of Mannheim

Page 2: Deliverable 2.2: Demonstrator on Multiscale Motion Estimation Florian Becker, Jing Yuan, Christoph Schnörr CVGPR group, University of Mannheim

Multiscale Filter Library

Implementation: ANSI C

Interface: C/C++ and Matlab (MEX)

Lowpass-filter: binomial filters

Resampling: spline interpolation (degree 2 to 5)

Scaling factor: selectable

Page 3: Deliverable 2.2: Demonstrator on Multiscale Motion Estimation Florian Becker, Jing Yuan, Christoph Schnörr CVGPR group, University of Mannheim

Multiscale Motion Estimation

Demonstrator 2.2

Implementation: Matlab

Multiscale framework for motion estimation:

• dyadic image pyramid → multiscale filter library

• underlying singlescale motion estimator → Lucas/Kanade

• image warping → spline interpolation

• warp rescaling → spline interpolation

• evaluation: synthetic PIV data

Page 4: Deliverable 2.2: Demonstrator on Multiscale Motion Estimation Florian Becker, Jing Yuan, Christoph Schnörr CVGPR group, University of Mannheim

Spline Interpolation: 360° Rotation in 23 Steps

original cubic spline

bicubicbilinear

Page 5: Deliverable 2.2: Demonstrator on Multiscale Motion Estimation Florian Becker, Jing Yuan, Christoph Schnörr CVGPR group, University of Mannheim

Spline Interpolation: 10 x Zoom

bilinear cubic spline

Page 6: Deliverable 2.2: Demonstrator on Multiscale Motion Estimation Florian Becker, Jing Yuan, Christoph Schnörr CVGPR group, University of Mannheim

↓↓

↓ ↓

W

W

E +

scale down

scale down

scale down

scale down

scale up scale up

warp image

warp image

estimatewarp

joinwarps

← next coarser level next finer level →

Page 7: Deliverable 2.2: Demonstrator on Multiscale Motion Estimation Florian Becker, Jing Yuan, Christoph Schnörr CVGPR group, University of Mannheim

Regularisation of Local Flow Estimation

Replace data term in variational approaches

M and d: from multiscale Lucas/Kanade estimator

Definition: compact ASCII file format for data term

Example: Horn/Schunck with replaced data term

Page 8: Deliverable 2.2: Demonstrator on Multiscale Motion Estimation Florian Becker, Jing Yuan, Christoph Schnörr CVGPR group, University of Mannheim

PIV image

New Data Term

confidential measurement

local estimation

Page 9: Deliverable 2.2: Demonstrator on Multiscale Motion Estimation Florian Becker, Jing Yuan, Christoph Schnörr CVGPR group, University of Mannheim

Horn/Schunck with New Data Term

λ=0 λ=10-4

λ=10-2