I just checked out your Voodoo camera tracker video..just brilliant.
It's very close to what would be needed..could you import each image and line it up with the camera and point cloud too?

Hi Chris
..yes.
insight3d will generate and match automatic points to calibrate the cameras with, but you can export just user vertices if you use "Modelling/Triangulate user vertices" instead of Triangulate all or Trusted Vertices. You can export the Camera data as an individual .txt file too. Insight3D seems to work mostly but it is very buggy (crashes, won't open saved files etc). I haven't tried "Bundler + CMVS + PMVS2" yet but it looks like a much more robust (but complicated) solution but I'm pretty sure you could do the same thing with it.

I made a test of 3 awful images (just 3 images, various focal lengths, bushes in the way, nothing vertical or horizontal..some walls look true but don't be fooled, the ground this house is on has been subsiding for 130 years) just to see what insight3d did..it seems to have worked in spite of the obstructions. I placed quite few of my own points in and it matched enough to work with.

I tried to match photo these images in SU and they are a nightmare..which it why I chose them. Almost all the Photogrammetry/photomatch demo's I've seen are of nice square US office blocks or otherwise unobstructed large buildings, not many of those in London.

I have a zip of:

1. the original images,
2. the full point cloud
3. the user points I placed, the ones that it matched
4. the camera data file

if you want to look at them I'll send them Chris (10MB)?

Coloured point clouds in SU: well maybe, kind of?

I just made a test "cloud" of multi coloured text full stops (Tahoma) in SU; arrow hidden, hidden leader line..they sit there in space as "face me" squares that remain a constant size..and they can be any colour you like - but you can't snap to them (which may be good!). I would think that one could make a coloured point cloud with them, if one knew what one was doing (which sadly I don't). Maybe one could place a construction point at the center of the full stops to get a snap-to functionality.

Just found these too, fantastic. Watch the videos?
http://www.pgrammetry.com/forum/
http://meshlab.sourceforge.net/
http://blog.neonascent.net/archives/bundler-photogrammetry-package/
http://blog.neonascent.net/archives/photosynth-toolkit/
http://blog.neonascent.net/archives/cameraexport-photosynth-to-camera-projection-in-3ds-max/

OK, this looks exiting (catch that Chateau facade!!): (sorry for the long post)

http://opensourcephotogrammetry.blogspot.com/
http://grail.cs.washington.edu/software/pmvs/

I've been using the insight3d application for a few evenings with mixed results..it will make a point cloud OK (really well actually from a tightly shot set of images - shot 1-2m apart with lots of overlap) but it seems to fall over if you use large data sets.

I can get a BW point cloud into SU OK (very slow, it won't export the colour data in the txt file anyway and the vrml export does not seem to export any polygon faces - which would be great help in re-orientating the cloud in SU (?); if you can ever get the vrml into SU of course).

What Insight3D does really well (and this python app by the look) is create calibrated cameras and camera positions for each shot - it seems to do that really well. Exactly the job that drives you nuts if you use match-photo alone with 4/5/6 or more images. I use a Canon DSLR with an EFS 10-22mm for shots. It has almost zero barrel distortion at 17mm and I still have trouble matching more than 3/4 images with reasonable accuracy (even after running PTlens over them).

Aside from accuracy, which is basic at best, a big issue with match-photo is when you go round a corner and lose sight of the origin. You're left to guess the origin, nudge the handles a bit, guess again, nudge etc etc. (or start over and match 2 models together) You can get a result of sorts but its way off ideal. There has to be better way?!

...so what would be perfect is some very clever guy (not me for sure) could write a ruby that could:

1. Read camera data output from either insight3d or this "Bundler + CMVS + PMVS2" python app.

2. Load the point cloud (with an option to reduce it to say whatever% ; a manageable density) and set up all the camera positions for each image.

3. Set up match photo scenes from the images ready to start modelling.

It would be fantastic to be able to start photo-modeling with perfectly camera matched and scaled scenes, with a light point cloud to snap to. (coloured or not)

That should let us model much more complex scenes and measured surveys with much greater accuracy. We'd have a real low cost alternative to the very expensive commercial photogrammetry packages LiDAR scanners and SU plugs (which I won't mention)?

I couldn't write a ruby script to save my life but.. anyone think its possible?

If coloured construction points are impossible..how does this work?

http://www.pointools.com/pointools-plug-in-for-sketchup.php

I've been playing with insight3d..which is beta,(try it and you'll see. TIP: when its processing data..don't use other apps, playmusic etc..just be cool, leave it to run..it crashes a lot if you don't) but frankly brilliant..it really works with a good data set of consecutive images it can create very nice, tight point clouds of real world buildings. insight3d + Sketchup could be real game changer. So, anyone up for writing a ruby script to get coloured insight3d point clouds into SU?