Fisheye lenses - When is using OCamCalib a good idea ?

[matthie...@gmail.com]

Hi,

I'm trying to calibrate somewhat accurately a camera, so later on I can given a pose and altitude estimate figure out the coordinates related to a pixel.

I'm wondering at what sort of fov of lenses it starts being a good idea to use OCamCalib, versus using the calibration model (pinhole?) found in OpenCV [1]

For info I'm trying to calibrate a gopro hero3[2] which has a diagonal fov of 150 deg.

As a follow up question, what strategies (bigger pattern, more squares ?) would be good to try if the calibration error (RMS) is quite high ?

Thank you,
Matthieu.

[1] http://docs.opencv.org/doc/tutorials/calib3d/camera_calibration/camera_calibration.html
[2] http://gopro.com/support/articles/hero3-field-of-view-fov-information

[Davide Scaramuzza]

Dear Matthieu

thanks for this question. See my answer below.

On 12.01.2014 04:34, matthie...@gmail.com wrote:
> Hi,
>
> I'm trying to calibrate somewhat accurately a camera, so later on I can given a pose and altitude estimate figure out the coordinates related to a pixel.
>
> I'm wondering at what sort of fov of lenses it starts being a good idea to use OCamCalib, versus using the calibration model (pinhole?) found in OpenCV [1]

With catadioptric cameras ( cameras plus mirror) it only makes sense to
use my toolbox. For fisheye lenses, if the camera has > 120 degrees,
then it makes sense to use my toolbox. The Toolbox of Bouguet (i.e.,
Caltech) or OpenCV would not work with >120 FOV.

> For info I'm trying to calibrate a gopro hero3[2] which has a diagonal fov of 150 deg.

For GoPro I only use my toolbox, and it works very well. I tried with
the Bouguet one but it didn't work

> As a follow up question, what strategies (bigger pattern, more squares ?) would be good to try if the calibration error (RMS) is quite high ?

Usually I always print my pattern on a A4 or A3 paper and never had
problems. You need a bigger size only if the minimula focal distance is
too large. But typically all lense can focus an image at a few
centimeters. Using more squares will not improve anything. Just make
sure that the
1. chessboard is always totally visible in the whole image
2. The white board of the chessboad is visible (otherwise the automatic
corner extractions doesn't work)
3. Make sure that the chessboard does not appear too small in each image
(as a rule of the thumb not smaller than half of the image size)
4. Take 10-20 images of the chessboard such that they cover the whole
camera field of view
5. Change the orientation (tilt and roll) of the chessboard

Best
Davide

___________________________________

Prof. Dr. Davide Scaramuzza
Head of the Robotics and Perception Group: http://rpg.ifi.uzh.ch
University of Zurich,
Andreasstrasse 15, AND 2.28, Zurich, Switzerland
Office: +41 44 635 2409
___________________________________

[Matthieu Tourne]

Hi Davide, and thank you for the detailed answer!

Assuming that I can calibrate my camera correctly, now I can use cam2world to map a point to a coordinate on the unit sphere (centered around my camera ?) 

Knowing my altitude from the ground, do I just need to do a simple projection onto that vector to get back a point on the ground ?

[Davide Scaramuzza]

On 12.01.2014 20:01, Matthieu Tourne wrote:

Hi Davide, and thank you for the detailed answer!

Assuming that I can calibrate my camera correctly, now I can use cam2world to map a point to a coordinate on the unit sphere (centered around my camera ?)

yes

Knowing my altitude from the ground, do I just need to do a simple projection onto that vector to get back a point on the ground ?

yes, it's simple trigonometry
Best
Davide

[pivaf...@gmail.com]

Hi Davide, just a question: if I have the point on the unit sphere, I have to assume that my camera is in the center of the sphere? Or it is camera focus that I have to put in the center of the sphere?

Thank you!
Francesco