Given how other types of image (CAT scans for example) are built through the application of software techniques, I suppose this was inevitable. A (near) future game changer for consumer cameras?
Impressive, logic and actually simple. Yes the time is coming where physical limitations of shutter speed, refraction of light etc etc are going to be addressed by cyber code. There is NO end in where 'will it end'. Then in the next 15 years you will see the might of the NANO... The new expense in photography is already shifting towards software.
The biggest challenge is the variance of the point spread, across the frame, with wavelength of light, and with aperture. They have addressed this nicely. This is a significant step forward, IMO.
i wonder how they come up with a good number for total number of patches to calculate.
The technique looks good looking at the video, depending on how they deal with the information, and how the runs compare to the current standards - they could really be changing things.
i wonder how they come up with a good number for total number of patches to calculate.
The technique looks good looking at the video, depending on how they deal with the information, and how the runs compare to the current standards - they could really be changing things.
My guess is that basically any number would work, with improved results as you use more 'patches'.
i wonder how they come up with a good number for total number of patches to calculate.
I'm guessing that it is empirical, being a balance between the number of patches (more patches, higher accuracy) and the costs of increasing patch density. In the corners, it is relatively important to have a higher patch density, since the point spread function changes rapidly from place to place. In the center, not so much.
ETA: If computational power were no worry, this technique could in theory be used to model and correct any lens. You'd test the lens, create a representation of its point spread function as a function of position in the frame, aperture setting, focus distance, etc., and perform your deconvolution based on that function and the reported EXIF information.
i wonder how they come up with a good number for total number of patches to calculate.
I'm guessing that it is empirical, being a balance between the number of patches (more patches, higher accuracy) and the costs of increasing patch density. In the corners, it is relatively important to have a higher patch density, since the point spread function changes rapidly from place to place. In the center, not so much.
ETA: If computational power were no worry, this technique could in theory be used to model and correct any lens. You'd test the lens, create a representation of its point spread function as a function of position in the frame, aperture setting, focus distance, etc., and perform your deconvolution based on that function and the reported EXIF information.
i wonder how they come up with a good number for total number of patches to calculate.
I'm guessing that it is empirical, being a balance between the number of patches (more patches, higher accuracy) and the costs of increasing patch density. In the corners, it is relatively important to have a higher patch density, since the point spread function changes rapidly from place to place. In the center, not so much.
ETA: If computational power were no worry, this technique could in theory be used to model and correct any lens. You'd test the lens, create a representation of its point spread function as a function of position in the frame, aperture setting, focus distance, etc., and perform your deconvolution based on that function and the reported EXIF information.
Would you mind deconvoluting this post?
"Efficient primal dual forward backward splitting method": you've never experimented with this? Even inefficiently? I'd expect dancers to be especially adept at this, and gymnasts...
how would this technology improve expensive lenses?
It would help nice lenses too, but not by nearly a much. Nice lenses are so much closer to perfect that the errors are more to do with manufacturing variance than they are poor optics. To get the same improvement you would need to calculate the point spread functions uniquely for every copy of the lens.
Paul is correct that it would be fairly easy to do computationally. It's more than a lookup table, but could probably still be done real-time in an FPGA.
"Efficient primal dual forward backward splitting method": you've never experimented with this? Even inefficiently? I'd expect dancers to be especially adept at this, and gymnasts...
I could do the split!
Still can to a certain extent but not backward. Maybe can correct it with software
"Efficient primal dual forward backward splitting method": you've never experimented with this? Even inefficiently? I'd expect dancers to be especially adept at this, and gymnasts...
I could do the split!
Still can to a certain extent but not backward. Maybe can correct it with software
