Using FullFrame Fisheye to create 3D Model for Tunnel

[geoclip]

Greetings, I have a project to create a 3D model of a small tunnel.  The tunnel is a four metres square opening and is about forty metres long.  To visually document the tunnel, I will capture high resolution 360 degree panoramas in the tunnel at 10, 20 and 30 metres into the tunnel. 

As it is difficult to capture sufficient overlap for modeling such a large tunnel surface using a narrow field of view camera/lens, I would like to use a wide angle full frame 8mm fisheye lens with the 1.6 conversion, I believe is equivalent to about a 12mm lens.  I am using a Canon T2i camera and a Samyang 8mm fullframe fisheye lens.  The shape of the tunnel should provide a reasonable angle of incidence of each image to the tunnel surface.  I would like to take six plus photos along regular sections of the tunnel at perhaps 3metre intervals (I will have to check the sidelap).

I have experimented using a fullframe fisheye lens for 3D modeling with Agisoft for small outdoor scenes.  I change the Exif focal length in each image and then process with photoscan.  I have had mixed results often very noisy, but it models.  I have created various models using a 4.5 metre pole with the camera pointing more or less straight down and I just mask out the pole and myself.  I take enough overlapping images to map a small scene - usually a number of rows like very low level aerial flight lines over the scene surface.  I have also used the same process using video mode with similar results.  I extract the frames and seed the EXIF with the correct focal length. 

I read the user manual and it sounds like for the tunnel scenario, it may be worth trying a full lens calibration, using Agisoft lens, on the full frame fisheye and seeing if the results improve the 3D modeling.

Has anyone tried calibrating a fullframe fisheye and had any success with outdoor 3D scene modeling?

Any tips would be appreciated

Thank you,
Jim

[fpbv]

Hi Geoclip

Can you send 4 images to me to do a test?

[Wishgranter]

http://www.asprs.org/a/publications/proceedings/baltimore09/0049.pdf

and

http://www.isprs.org/proceedings/XXXVII/congress/5_pdf/134.pdf

and this one go for AGISOFT team 

http://upetd.up.ac.za/thesis/available/etd-01232009-161525/unrestricted/dissertation.pdf

[geoclip]

Thank you for the great links to the fisheye calibration papers.  I will read them carefully.  I did a camera/lens calibration last night.  I will try a few test images and see if I can come up with a solution that models well.  Thank you again for your replies - much appreciated.

Jim

[EricM]

If you find a good workflow, please let us know. I too have a full frame fisheye lens and would love to use it in tight spaces, but these papers are a bit over my head...

[RalfH]

I don't quite understand why you are so keen on using a fisheye lens. The point of this appears to be to be able to capture large spaces with only a few images. But memory is cheap, cameras are fast, and the larger the area on an individual image the poorer its spacial resolution. I would always prefer using a standard or moderate wide angle lens. Less distortion, higher resolution and no hassle when trying to create models. Just take more pictures.

[geoclip]

Hi Ralph,  I agree it is easy to use a normal to moderate wide angle lens for most cases.  I think there are special cases where it would be much easier to use a fisheye lens such as this case being a tunnel.  The links to the papers kindly noted above use fisheye lens for narrow alleys, etc.  It would take hundreds of photos to cover the surface of the tunnel having sufficient overlap using a normal FOV lens.  Also, these are not just hand held snap photos - a tripod was needed.

[RalfH]

Yes, I can see the point about taking hundreds of photos with a tripod.

As for the field of view, the problem with PhotoScan (and with any comparable 3D reconstruction approach) is not so much the distortion by the fisheye lens (which can be corrected by lens calibration) but the large changes in viewing angle from image to image. Consider a normal lens: the FOV is relatively small, so the camera positions have to be relatively closely spaced, which in turn means that the viewing angle from a camera to a given object does not change very much from image to image - if the object is looked at from a similar direction, it can easily be recognised by an algorithm. Now consider a fisheye lens: the FOV is large, the camera positions can be far apart, which in turn means that the viewing angle changes drastically - now, if you look at the same object from very different directions, it will be much more difficult (and error-prone) for an algorithm to correctly recognise the same point in consecutive images. This is what I see as the biggest problem with fisheye lenses, and this is true no matter how well your lens is calibrated and how well the camera is modeled in the 3D reconstruction process.

P.S.: You can ponder this problem while trying to manually find corresponding points in the tunnel image set. A certain degree of self-similarity of the tunnel wall surface makes it even more difficult.