Blur-Kernel Estimation from Spectral Irregularities
ECCV 2012
Amit Goldstein and Raanan Fattal
Results Page
Page contains additional tests over real and synthetic images.
Instructions: left-click on each image to view in full resolution.
Extracted blur kernels are shown on top of each image.
Real Images (aquired with motion blur)
Venice
Torres Del Paine
Torres Del Paine
Real world images taken from Cho et al. 11'
Synthetic Examples
Images below are some of the synthetic images used in our dataset for the quantitative evaluation
Failure cases
Blurred
|
Cho,Lee 09'
|
Cho et al. 11'
|
Our
|
To further investigate this failure case we obtained another image of the same bridge
("Zakim" in Boston) and computed its autocorrelation function.
The repeating structure in the image (the bridge wires) translated
into severe biases to this statistic (and ones at the scale of the blur; larger-scale
effects would have not been a major problem).
These biases undermine the ability to whiten the image spectrum and
estimate the kernel's power spectrum function.
Similar sharp image
|
Auto Correlation of sharp image
|
Significant spatial variability of the blur is not compatible with our model assumptions
Significant variability in focal blur is not compatible with our model assumptions