Demosaicing on Fujifilm X-Trans Sensors | Hacker News Reader

Demosaicing on Fujifilm X-Trans Sensors

public.asu.edu

65 pointsenthdegree10 comments

Interesting. I've seen some other work done on gradient-aware demosaic using directional masks, and the devil appears to be in the details: it can reconstruct a lot of sharpness when images are aligned with masks, but it also tends to take images that aren't aligned with the masks and "forceably align" them -- e.g., circles get turned into polygons, etcetera.

Perhaps the additional green resolution on X-Trans sensors can help avoid this some. But I'd be interested to see how well this ends up doing on trickier images...

I'm looking forward to any improvements he can achieve. I think x-trans was overall a bad idea from Fuji.

Higher megapixel sensors have allowed the removal of the low pass filter in traditional Bayer sensors, without all the post-processing troubles caused by x-trans.

I love my fuji. But x-trans processing is a real pain point, so this is cool. Case in point; google photos won't process any RAF files uploaded. Someday, I assume. But, not ideal for someone who shoots raw.

I have an X-100s. I really like the physical design of the camera. The fact that they did firmware updates that gave new features was also great.

However, I shoot RAW and x-trans seems to be an issue. I noticed quite a few of my photos have weird color issues. Objectively my old, cheap Nikon DLSR was producing better photos.

Adobe products are known to have bad X-Trans processing. There are tons of debates about this on Fuji forums. Other products handle X-Trans better like RawTherapee, Iridient and Capture One.

I usually use Darktable. RawTherapee, Lightroom and even the in camera processing seem to have color issues with certain photos. Maybe it is not xtrans but my old Nikon DSLR and my new d750 don't have these issues.

Interesting. Never heard about that. The typical complaints are about sharpening, especially foliage.

I bet they're trading off detail for color accuracy. If only we knew how they made that tradeoff...

someone is wasting a phd scholarship to solve a problem that only exist because people keep dumping money on a company that damages their own product by now releasing source or specs?

OP here, you're right and it's even worse than you say. After my brief literature review I'm starting to suspect that the entire problem of demosaicing is ill-posed in most of the ways it is studied (certainly including this little side project). Some of the survey papers I saw talk about this. The color image pipeline is basically ignored in all the analysis I've seen, and to the best of my knowledge the common assumptions are based on dubious heuristics.

On the other hand it works well enough...

Reply on news.ycombinator.com