[GRASS-user] Re: [GRASS-dev] DEM creation from stereo pairs

[Christoph Högl]

Am Samstag, 12. Dezember 2009 22:18:37 schrieb Hamish:
> Hamish:
> > > have a look at:
> > >  http://grass.osgeo.org/gdp/stereo-grass/
> > >
> > > I'm pretty sure there was a FOSS conference paper on that
> > > method a few years ago.
> 
> Michael:
> > Unfortunately, the requisite "stereo" package not only has
> > not been maintained since 1997 (according to the stereo-grass
> > page), but its web page has disappeared too.
> 
> the question is, how well did it work back then? working code
> remains working code. :)
> 
> you might try to look for a copy of the website on archive.org,
> if not maybe I or one of us has an old copy of the source
> somewhere in the old dusty attic.
> 
> 
> shrug
> Hamish
> 
Well, the package did work and sometimes still does.
it's job.

The accuracy of the process after tweaking it a bit for small rounding mishaps
is about 12 to 14 bits (16bits data input). In comparison a hugin-based 
process achieves a little -literally- 1 bit more of accuracy at about 5 times 
the amount of time for data preparation (human) and 3-fold data processing 
time (cpu-time).

In the old times of stereo I even had to reduce resolution of the source 
images in order to get them processed with 256M of RAM (8 Mega-Pixels 
(3344x2508) down to 4 MP (2364x1772) at that time)

It still provided about pixel accuracy with regards to optical problems 
(mirage/thermal inversion/turbulence of the air caused by heat/cold).

To make it more feasible: 
Using a mounted camera 8 MP (which is common these days in this work field), 
provides at a field of view of 45° a angular distance of about 48 arc seconds 
or about 1/75 of a degree which gives a accuracy of about 23cm at a distance 
of 1000 meters. The atmospheric effect at that distance is way beyond (~ 3 
meters)
If you are after taking photos in less ambitious way, let's say shooting an 
excavation from 100m or less, the accuracy is -> 2.3cm at best, obviously 
better near the center of field (less lens distortion, ...) and nearer to the 
lens (less atmospheric (d)ef(f)ects).

Even if we take a sefety measure of factor 5 for each photo taken using a 
shaky tripod or even factor 10 from a kite mounted camera and multiply all 
worse effects instead of flattening out, the accuracy still exceeds most needs 
at archeological sites (2.54cm = 1 inch) at distances of up to 25 m by a 
factor of 8.

The worst case i had with stereo was a misplacement by 1.5m  at 100m distance 
using scans of ballon-mounted photograph from the late 30ies combined with
a 2003 photo taken using 4MP camera. (BTW, the 4MP was the culprit in the end)

One of the miracles using stereo was the ability to revisit a place in the 
desert by triangulation using stereo after 8 years with 30cm accuracy and 
pretty unsharp joshua-trees in the background (at distances of ~15m and ~25m  
unsharp, 6 team photos taken in front of a car, taking the rim as rule using a 
4MP camera). GPS Data was misleading by 4-8 meters. 

The main reason for using stereo is it's ability to combine different sorts of 
source materials (plates/film/digital imagery) in pretty short time. (I use it 
still as stated above for small projects, where only a few points in 3d-space 
need to be 'surveyed')
As stereo takes more than 2 pictures it is even more powerful than most 
standard off-the-shelf commercial solutions.

The main reasons against it's use is the lack of automation in the process, 
the lack of dedicated support (I'm not able to provide it, though stereo would 
really deserve it) as well as the lack of camera/lens adjustment combined with 
a really awful codebase.


Hope that helps a bit,
Christoph