0
$\begingroup$

For a couple of years now, I've been playing with trying to create scenes from photos (I've found very few references on this; if you know of any, please let me know). I've run into an odd problem, though, with some photos of the 1904 Worlds Fair: I can't come up with any camera orientation or position that matches the photo to the supposed positions of objects in the scene.

The blend file at the link below demonstrates the problem. Here's a screenshot:

Blender screenshot showing vanishing lines, reference photo, and ground features

http://pasteall.org/blend/index.php?id=44609

The camera contains the highlighted mesh as a child. The mesh represents the image plane and contains the vanishing lines, which I use to orient the camera. The mesh also has an image "empty" as a child, which contains the reference photo. The photo is from the wireless tower's 100-foot platform.

The colored polygons represent ground features, traced from this map (look for "Plaza of Orleans" on the left side):

http://atthefair.homestead.com/Misc/Maps.html

I just can't get the camera to line up. For example, I can't get the two strips of grass in the middle of the photo to align with the corresponding polygons. The camera should be at roughly the right position, but even if I put the camera wherever I want and tweak the orientation, I can't get it to line up. Although the original map isn't perfect, my research indicates that it should be pretty close.

I'm wondering if there may be some distortion in the photo. But the lines all look straight, and the sphere on the railing looks circular, so it would have to be a distortion that doesn't affect these.

So, if anyone can figure out an alignment that works (and preserves the vanishing points), or knows of some possible distortion in the photo that's preventing it, I'd appreciate any help.

[Update 1] I installed BLAM and gave it a try. The angles were similar to what I was getting. Things still wouldn't align. Of course, this isn't the best subject for vanishing point based analysis - there just aren't many long parallel features.

I would consider the map being in error, but I compared it with satellite images (using what few artifacts are left from the fair), and it appears to be pretty accurate overall.

Is there any sort of photographic distortion that would only affect the depth? I can't conceive of one, but I'm no optics expert.

$\endgroup$
  • $\begingroup$ Check this: blender.stackexchange.com/questions/33625/… $\endgroup$ – Denis Dec 9 '16 at 4:06
  • $\begingroup$ Oh great :(. Don't tell me someone else has done what I've been slaving away at. I swear I've searched for things like this many times and never ran across any for Blender. I'll have to take a closer look at it over the weekend. $\endgroup$ – Jabberwock Dec 9 '16 at 4:13
  • $\begingroup$ You have three variables: the lens (or angle of coverage), the placement of the camera and the dimensions of the object. Without knowing at least two of them is hard to match. For the lens try using Blam, to place the camera correctly lock the camera to view and place the camera that way (read: blender.stackexchange.com/questions/40308/…) . $\endgroup$ – cegaton Dec 9 '16 at 5:43
  • $\begingroup$ The "don't tell me" part was a reference to the work I've put into the scripts to orient the camera based on the vanishing points, which looks like what BLAM does. And the fact that all of the searches I did never turned it up. I'll have to look closely at it and see how it works under the hood. As for variables, I do have three vanishing points (to some level of accuracy), which gives me the camera orientation and focal length. The camera position, particularly the distance to/size of the object is ambiguous from the photo, but that's where the map comes in. I should have all I need. $\endgroup$ – Jabberwock Dec 9 '16 at 17:35
  • $\begingroup$ After spending the weekend learning more about linear algebra and matrices than I ever learned in college, I have an idea. I found a very good paper - 3D Reconstruction from Multiple Images by T. Moons et al - that mentions a skew parameter, which represents the image plane not being perpendicular to the camera's optical access. I could see this at least partially causing the problems I'm seeing, although I'm not sure it's realistic for the skew to be bad enough to account for all of it (on the order of 10 degrees or more). Is anyone familiar with the effect of skew? $\endgroup$ – Jabberwock Dec 12 '16 at 20:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.