radiometric alignment and vignetting...
TRANSCRIPT
Radiometric alignmentand vignetting calibration
Pablo d'AngeloUniversity of Bielefeld
Pablo d'Angelo CCMVS 2007
Overview
● Motivation
● Image formation
● Vignetting and exposure estimation
● Results
● Summary
Pablo d'Angelo CCMVS 2007
Motivation
● Determination of vignetting and exposure● Traditional approach
– Flatfield + camera response calibration inthe lab
● Simpler: use overlapping images
● Applications– Spatial & photometric mosaicing– Recovery of scene radiance, required
for radiometric reconstruction methods(Shape from Shading, Photoconsitency)
Pablo d'Angelo CCMVS 2007
● Image irradiance:
● Natural vignetting:
Image formation
Radiance L in [ Wsrm2 ] Irradiance I in [ Wm2 ]
I= k 2 M L k aperture value
M vignetting function
M=cos4
Pablo d'Angelo CCMVS 2007
● Integration on sensor ● Used model
Image formation
Radiance L[ Wsrm2 ] Irradiance I [ Wm2 ]
in image plane
Camera responsef wi eML
e=st k 2 effective exposure
M vignetting functionwi white balance factori channel numberk aperture values camera sensitivityt integration time
Pixel values Biin [colour levels]
Bi= f w i eML
IntegrationstI
Pablo d'Angelo CCMVS 2007
Natural vignetting
● Light falloff due to– Apparent area of exit pupil, cos(b)– Larger effective area on film, cos(b)– Larger distance to image corner,
cos2(b)
● Natural vignetting not reduced by stopping down
● Lenses are complicated,cos4 does not apply for many designs
Senso
r
© Paul van Walree
Pablo d'Angelo CCMVS 2007
Optical vignetting
● Entrance pupil shaded by lens barrel
● Optical vignetting depends on aperture
© Paul van Walree
f 1.4 f 5.6
Pablo d'Angelo CCMVS 2007
Mechanical vignetting
● Light is blocked by lens hood, or other objects
● Results in unrecoverablevignetting
© Paul van Walree
Pablo d'Angelo CCMVS 2007
Vignetting of a real lensFlatfield image● Image of a homogeneous white surface
● Observation– Radial falloff– Not centred
Image by Goldman and Chen (2005)Contrast enhanced and quantized for better visibility on a beamer
Pablo d'Angelo CCMVS 2007
Modelling vignetting
● Non-Parametric models– Flatfield image
● accurate● acquisition cumbersome
● Parametric models– Radial polynomial model
– Allow center shift c
M=1 r62 r
43 r21
rela
tive
irrad
ianc
e
Pablo d'Angelo CCMVS 2007
Modelling thecamera response● Non-Parametric
– Can model all shapes– Dense data required
● Parametric– Polynomial
● Less parameters● Dense data required
– PCA basis of 201 response functions (Grossberg, Nayar)● First 3 Eigenresponses describe all 201 functions well.● Strong model, can be used with sparse data
Pablo d'Angelo CCMVS 2007
Grey value transfer function
● Scene points imaged in registeredimages
● Joint histogram– cannot model vignetting
● Gray value transfer function
L1=L2
B1=B2= f e1M x1 f
−1 B2e2M x2
B1
B2
Pablo d'Angelo CCMVS 2007
Estimation usingcorresponding points● Estimation of the transfer function by minimising
● Other approaches (Goldman 05)– Alternating minimisation steps of
for L and e,M,f– Large number of variables– Slow convergence
et=d B1−B2
e=∑ j d B j− f eM L j
d distance metric
L1=L2
Pablo d'Angelo CCMVS 2007
Estimation usingcorresponding points● Estimation of the transfer function by minimising
● Symmetric transfer error:
● Minimising es using a set of (B
1,B2) measurements
yields– Exposure, vignetting, white balance– Camera response
et=d B1−B2
es=d B1−B2d B2−B1
d distance metric
Pablo d'Angelo CCMVS 2007
The exponential ambiguity
● Simultaneous estimation of exposure and camera response subject to exponential ambiguity
● Calibration and Radiometry– Either camera response or exposure need to be known
● Removal of exposure, vignetting and WBdifferences– Find any solution, correct images – Use inverse response to transform into original grey
value space
Pablo d'Angelo CCMVS 2007
Extraction ofcorresponding points● Avoid outliers due to misregistration
– Extract points in low gradient areas● Need points at all radii r
– Bin points by distance from image centre– Select points with lowest gradients
allcorrespondences
extractedcorrespondences
B1
B2
B1
B2
I 1 I 2
Pablo d'Angelo CCMVS 2007
EvaluationSynthetic example
● Simulated panorama with 6 overlapping images– Gaussian noise (2 grey values), quantisation, outliers– Estimate vignetting and exposure,
analyse error to ground truth
300 points
10% outliers
Pablo d'Angelo CCMVS 2007
VeniceVignetting correction
Pablo d'Angelo CCMVS 2007
Green LakeVignetting, exposure and WB correction
Original images
Corrected images
Pablo d'Angelo CCMVS 2007
Green LakeResults Goldman & Chen
proposed method
Goldman & Chen
Pablo d'Angelo CCMVS 2007
Spherical panoramaOriginal images
Pablo d'Angelo CCMVS 2007
Spherical panoramaVignetting, exposure and WB correction
Pablo d'Angelo CCMVS 2007
Summary and conclusion
● Estimation of radiometric camera parameters from overlapping images– exposure, vignetting, response curve, white balance– Robust approach, better results than previous methods– Compact parametrisation, fast estimation
● Applications– Radiometric calibration, estimation of scene radiance– Removal of photometric distortions from image
mosaics
● Implemented in Hugin, http://hugin.sf.net
Pablo d'Angelo CCMVS 2007
Radiometric alignmentand vignetting calibration
by
Pablo d'Angelo
Pablo d'Angelo CCMVS 2007
EvaluationImage blending