The girl with the dog
Image description: When our colleague Sarah heard that we are doing deblur research, she sent this photo to us and say: "try this one!" And we were hammered. You can see why below.
To estimate the blur kernel, the algorithm will automatically find some edges that it thinks are "good" or "reliable", based on some criteria. In this case, even there are some lines in the background, the algorithm does not think it can trust those lines, and eventually decide that the image is not blurry. That's why the output kernel is so close to a dot. And, applying this kernel to the image won't change the content much. However, if you are interested in pixel-by-pixel difference, download the full-res images first :)
We agree that the algorithm does not always work well, and this is the case it fails right now. As we improve the algorithm, we hope we can get better result later on. In the meanwhile if you have an algorithm that works better on this example and you'd like to share it with us, we'd like to learn from you.
Buildings Far Away (G. Botet, France)
Image description: The first contribution we received from users. He/she wrote: "I took it yesterday morning from the window of my kitchen. Hence it is not famous (yet), it just for the technical interest : I had a 150mm tele lens on my Canon APS DSLR, no stabilizer, and 1/30 s speed (speed was too low indeed ;-). Such things can happen to a beginner.. ".
This one is interesting to us since it was shot by a tele lens, which we never tried. The deblur result is reasonable, especially for the buildings. Of course the output image quality is still not super, I think it can be further improved using other tools.
Botet was kind enough to provide another clean shot of the same scene, which is shown at the bottom. There is no magic in deblurring, it is a reverse process and can always lead to various artifacts. So if you are serious about photographing, try to take a clean shot everytime!
Cairo 1979 (Jeremy Tregler, Downers Grove, IL)
Image description: This one was contribued by Jeremy from illinois. He wrote: "It was a photo of my father touring Cairo during his time on the USS Saratoga in 1979. It looks as though the photographer moved the camera at the time of the photo, but I can't tell of that's the only damage to the image. "
This is another tough example as the image was scanned from an old photo. We managed to restore a sharper, cleaner version. What is interesting here is that we found the image was not blurred uniformly. Since our system only deals with uniform blur, we estimated kernels from four different parts of the image (see bottom), and then finally blended the four deblurred verions together to create the final result. This is a practical workflow to deal with non-uniform blur using the system. From the result it seems the original shot was not focused very well on the foreground, so there is also a litte bit out-of-focus blur on the faces. The camel is a big winner here :)