Hi, we develop software that takes portrait photos from a digital
camera. Now the camera is a bit too close so the face looks a bit too
rounded. The optical solution is to move the camera away from the face
and zoom in a bit, but we want to do it in software.

We have tried the 3rd order polynomial method to correct radial lens
distortion, which is mostly used to fix barrel distortion at the edges
of the image, but can be adjusted to modify the center of the image
which is what we are mostly interested in.

My question is, is there any standard algorithms that can correct this
effect better? Even better would be ones that we can give known
parameters too (like distance, lens size etc) and get a fairly
definitive correction formula, rather than just playing around with
parameters until we find something that looks right.

Hi, we develop software that takes portrait photos from a digital
camera. Now the camera is a bit too close so the face looks a bit too
rounded. The optical solution is to move the camera away from the face
and zoom in a bit, but we want to do it in software.

Oolis Kraprin wrote:
Hi, we develop software that takes portrait photos from a digital
camera. Now the camera is a bit too close so the face looks a bit too
rounded. The optical solution is to move the camera away from the face
and zoom in a bit, but we want to do it in software.

Am I right in thinking that your problem
arises because the face is stuck on the
front of a round head?

That is to say, in
an extreme case the camera will not be able
to see any ears because it can't see around
far enough.

If so then your solution is to convert the
2-D face into 3-D by modeling a round head.
Strictly speaking, you don't have enough
information to do this, but heads have a
standard shape so you can make assumptions.

If so then your solution is to convert the
2-D face into 3-D by modeling a round head.
Strictly speaking, you don't have enough
information to do this, but heads have a
standard shape so you can make assumptions.

Well we do take images from all sides of the face, but creating a 3d
model is not the solution for us as we are primary concernd with the
2d representation of a 3d object. Its a lot like projections in world
maps really, there is no perfect way to represent a globe on a flat
piece of paper so you need some model to represent the 2d effect. I
guess you are right in a way, the best way to represent the planet is
using a 3d globe. However the best way to represent a planet, or
someones head, in 2d form can only ever be some projection of the 3d
solution.

It appears using macro mode on the camera helps a bit. But our 3rd
order polynomial works too, albeit in a trial and error way.

Oolis Kraprin <Oolis.Kraprin at (no spam) googlemail.com> writes:
....
It appears using macro mode on the camera helps a bit. But our 3rd
order polynomial works too, albeit in a trial and error way.

Just a quick comment: I don't think Martin Leese meant that you should
be generating 3d representations as finished output. I think he was
suggesting that you do the following:

- Get the picture from the camera

- Using info about relative location of camera + head and also some
assumptions about the shape of a human head, calculate what "bit" of
the head each pixel in your image represents. This is going to give
you some texture map on the surface that approximates the head.

- Now find a transform that is "rendering" the textured head, from
further away with a saner focal length.