In the 3D-Pop Flickr group you have to browse a bit beyond that George guy that is all over the place. This one is very convincing to me.
I agree with SimpleJoy that primeshooter is the champ in that group.
Bas
Having worked my way through those, I'm more convinced than ever that the trick to this 3D look is to have a primary sharp subject in the foreground against a
slightly out of focus, but still recognisable, background. Just the right amount of background blur is essential. Too little and it doesn't work, and if the background is too blurred, the image flattens.
That's what I expected the result to be when I started my first little study of 3D pop in 2014. People didn't generally agree on this, which is a pitty because it's not terrible to enhance this attribute in computational postprocessing. The only thing they did agree on is
Chromostereopsis. Of course, lots of images with high sharpness and a gentle transition to not-too-out-of-focus also happen to benefit from chromostereopsis -- especially since human subjects are reddish.
Incidentally, in every study I've done where I had humans evaluate image quality, I've found that
humans are remarkably inconsistent about identifying good/bad attributes. The first was having humans evaluate the quality of repairs of the
Fuji X10 white orbs defect. The DPReview article linked explains how Fuji's modified sensor repaired the white orbs problem, but it actually didn't repair it. The horrific blooming is still there, but slightly less and Fuji stopped insanely sharpening the edges of orbs. Compare the images in the DPReview article to the computational repair done by my free DeOrbIt software working with images captured using the original sensor; here's a slide from my Electronic Imaging 2013 research presentation on DeOrbIt:
It's a highly credible repair to say the least, and looks much better than the "fixed" Fuji sensor produces. In case you're wondering how it works, DeOrbIt (1) identifies the white orbs, (2) uses an inpainting algorithm to replace each orb with credible background, and finally (3) applies computational relighting to add natural-looking smooth highlights.
However, not only did
DPReview staff accept the very low-quality repair of Fuji's "fixed" sensor, but when (before Fuji made their repair) I gave people the ability to run their own images through a web form version of DeOrbIt and score the repair quality, DeOrbIt got a mix of very high and very low scores. Why? Well,
the low scores mostly came from repair of user images that did not have the white orbs defect! People submitted shots with bokeh or other image content that looked a little like white orbs, and deorbit quite correctly recognized that those were not the defect and left the images untouched, causing the submitters to give very poor repair quality scores... In fact, online, a large fraction of people's images posted online supposedly showing how bad their camera's white orbs defect is are actually images that don't have the defect at all.
I see the same human inability to recognize image properties every time I teach my "cameras as computing systems" course. Students have trouble correctly identifying the various types of artifacts caused by lenses, sensors, etc., even after being taught how to recognize each.
