micahmedia: For the full story open the resolution charts from Imaging Resource. The color star targets still look horrid.
This is still not ready for prime time.
Kudos to Adobe for taking up the challenge though! I'm curious if they can still take this further. Fuji would be foolish not to share some their secrets with them, but...well, I suspect they -are- foolish.
The new ACR demosaicing is based on the algorithm that Fujifilm gave to Adobe (and Apple). However, demosaicing is just the first step in RAW processing.
beckmarc: Great news on the Fuji x trans sensor support
It is supposed to use the improved demosaicing algorithm that Fuji gave to Adobe. That hopefully means less color bleeding and improved detail with foliage.
Apparently the Express version doesn't have local adjustments and dust removal features.
Marcin 3M: I wonder if the shake trajectory was extracted from the image, or rather recorded by accelerometers. They stated that it was calculated, but this calculation is time-consuming. I would like to know the theory behind.
My bad. Convolution is actually algorithmically more complex than FFT. So yeah it's almost certainly impossible to do anything better than O(N * log N).
It might not be necessary to FFT the whole image. They might be able to get away with FFTing just certain fragments of it. Once the kernel is known, of course the deconvolution must be run on the full resolution.
I haven't done deblurring, but have done deskew / auto-straightening of scanned images. First you find a usable area that you analyze. Or let's suppose you do Hough transform to find an angle of a black border. Instead of computing the sine waves at full resolution, you do a low resolution estimate first, then once you have the approximate angle, you compute it at higher resolution, but only around the estimated angle.
I'm just saying there might be ways to avoid having to FFT the entire image.