Hacker News new | ask | show | jobs
by Anotheroneagain 694 days ago
You can get a rectilinear lens instead, you know.
5 comments

Optimising for low distortion means trading off against something else - sharpness, brightness, size, weight etc. Smartphone cameras have become so good because they're very intelligently optimised using a hybrid of hardware and software.

DSLR/mirrorless users still use lens correction (either in-camera or as part of the post-processing pipeline) because even a big, heavy, expensive pro-quality lens is still imperfect in ways that are relatively easy to compensate for in software.

https://www.canon-europe.com/pro/infobank/in-camera-lens-cor...

This isn't about correcting for minor imperfections, but converting the image from the wrong kind of lens.

See https://m.youtube.com/watch?v=6toiNmZ_e4I for the difference.

Sometimes computer vision applications require rectilinear images, but you don’t have a chance to choose the hardware, or it was chosen with other constraints in mind. No reason to dump on someone doing research to rectify an image in a novel way.
Sometimes you can, yes, if you are picking the lens with which a subject will be photographed -- you can get down as low as 9mm on 135-film area, and still buy a relatively rectilinear lens.

Sometimes you can't get a rectilinear lens, though: If I want to shoot wide angle on my phone, curvilinear will have to do.

Sometimes you don't even have a lens, you've just got a photo, and that photo is curvilinear.

Novel ways to adjust for distortion are always nice to have in the toolkit.

Sometimes you don't even have a photo, but rather a synthesized image that looks as if a lens was in use.

Sometimes, you want to have a system that's able to self-correct after you slap a random lens on it as it's working. Or, you're looking through a translucent material, which acts as an ad-hoc lens with unknown parameters, and you want to compensate on the go.

Point being, methods of on-line calibration without use of special calibration setup (like ArUco boards mentioned elsewhere) have a wide range of use cases and are always welcome.

I know very few 35mm format lenses with NO distortion.

The two I know of with the least distortion are actually primes from the 1980s. Nikon began allowing a small amount of distortion in their new prime designs circa 2010, choosing to correct it with an in-camera profile.

It's not as bad as it sounds. Getting rid of that last bit of distortion may require relatively major tradeoffs in other areas like brightness.

There's a number of lenses which prioritize distortion correction because they don't get to have lens profiles. Though even low distortion wide angle lenses generally retain low levels of high order distortion (i.e. straight lines become slightly wavy across the image, instead of having a large amount of low-order distortion, i.e. being simply bent strongly one way or another), see e.g. Laowa Zero-D lenses.

I do actually think the OEM design approach is better overall. It's a lot easier to near-perfectly correct high amounts of low order distortion than it is to make lines with a slight amount of 6th? 8th? order distortion actually straight. Even if the resulting raw image of the OEM lens looks more like a fisheye than a rectilinear lens.

Would one og these be the rectilinear 15mm?

Great lens - and a huge front element…

I used that in the 90’ies to make QTVR with a (really expensive) Kodak DCS 460 - 6MP on a DX sized CCD no display and the large 340MB PCMCIA disk ;-)

Zeiss cinema lenses (in particular master primes) have the least distortion I’ve come across
and a spherical cow while you're at it.
They're commonly available, not something exotic.
Are they? How do you store spherical cows? Do you need to roll them out to pasture and then back into the barn, or do they roll on their own?
I mean you have to pick between fisheye and rectilinear lenses when you buy a wide angle lens. This is completely unnecessary, you only need to pick the lens that you actually want.

Why does everybody doing as if I propose something outrageous or impossible?

> This is completely unnecessary, you only need to pick the lens that you actually want.

It sounds like you are thinking only in the context of photography. In robotics and machine vision applications you often choose the fisheye lens because they are cheaper than the rectilinear lenses with the same FOV. (if a rectilinear lens is even available in the form factor and FOV you need.)

So what people do in those situations is that they get a crappier lens and they calibrate it so the algorithms know how much to correct for its crappyness. That is where this kind of calibration really shines.

I think the down-voting was harsh, it usually gets corrected in no time.

That said people here are interested in the different ways of solving a problem, if not for anything else but to tickle one intellectually. So yes rectilinear lenses exist, but that does not mean that computational methods are uninteresting or useless. For one thing, one need not purchase different kinds of lenses.

These aren’t even fisheye examples.
Even the best and most expensive professional lenses have some barrel and or pincushion distortion.

It’s unreasonable to expect never to need any correction, and it’s actually a really interesting, non-trivial problem, to tinker with.