We take a ground truth image and blur it using a synthetic camera
motion. We then recover the deblurred result using handpicked
deconvolved patches as initialization and using the ground truth image
as initialization. We compare them with the deblurred result of the
spatially invariant blind deconvolution technique proposed by Shan et
al [1]. These results are discussed in Section 6.1 of our paper.
Inferences from these tests:
6 degrees of motion can be
reasonably approximated by 3 degrees
of motion (roll and x,y-translations) and,
With a perfect
initialization, we are able to recover a very good
deblurred result. Hence better spatially invariant deblurring
techniques can be plugged into our system for better results.
