On 6/23/13 7:21 AM, Tac Tacelosky wrote:
I've been using our streetviews (customstreetviews.com) to update OSM,
and would like to throw a few ideas out for consideration.
Street-level photos have a lat/long, but that location is not the
location of the item in the photo. Panoramic images, like ours, have
a lat/long of the camera, but when looking at a flat, normal view of
an image, we also store the heading, field of view (aka zoom), pitch
and size of the photo. Obviously, all those things can be put into a
single url. I realize at this point we're just talking about tagging
an object in the database with an image url, but wanted to bring this
idea up.
i spent some time working as a camera modeller at GE Research
in the late 80s/early 90s and can contribute some input into the
way that cameras are normally represented. while my work at the
time was primarily with aerial and satellite photography, basic
principles are basic principles.
we are making pinhole camera assumptions here. pretty much any
camera that OSM mappers are likely to encounter will easily be
modelled with pinhole camera assumptions. things that aren't
necessarily pinhole include SAR imagery from aircraft runs
(side looking along the flight path) and satellite imagery (frame
lines taken as the satellite progresses on its orbit.)
given pinhole assumptions, a camera can be modeled with 7
parameters:
x,y,z - translation within the same coordinate system as the
objects in the image. lat, long, and elevation would do for this
the three Euler angles, omega, phi, kappa - tip and tilt of the
camera, describing how it is aimed. 0,0,0 imply the camera is
aimed straight down aligned with the x & y axis, and the
angles are applied in order to position the sight line of the
camera.
http://en.wikipedia.org/wiki/Euler_angles
focal length - expressed as a ratio of the zoom and the
negative size. note that just knowing the lens used (55mm,
35mm, whatever) is not useful unless you also know the size
of the image sensor.
however, the 7 parameters rarely produce an accurate mapping
without some processing. in the stuff i did, the euler angles
were usually known pretty well (gyros do that decently) but
frequently lateral translations were off noticeably. focal lengths
were generally very reliable.
so you have to do something to adjust the 7 parameters.
this is normally done by identifying specific points in the
image that map to known points in 3D space, and running
a resection algorithm of some kind. i always used a USGS
algorithm that was public domain and converged very
quickly. haven't played with street level stuff much.
On a related note, I think it would be valuable to have all of the
available photos in a single, separate database, which could be
referenced by OSM objects. Something like commons.wikipedia.org.
nice idea but can you say "herding cats"?
richard
_______________________________________________
Tagging mailing list
Tagging@openstreetmap.org
http://lists.openstreetmap.org/listinfo/tagging
