View RSS Feed

Redrobes blog about playing with the technical side of graphics.


Rate this Entry
by , 12-01-2010 at 02:44 PM (9369 Views)
So this Insight3D is a gem of an app. You can find it here:

As stated on the page "it hopes to be "Photosynth + Canoma". Or "boujou + Imagemodeler". But it's also free and opensource (GNU AGPL 3)" - ok thats me sold then... Its not quite up to the mentioned apps but its heading in the right direction and seemed very stable to me. At this point its not super user interface easy but once you have run it through a few times then its not so bad. I think if they reordered the buttons to layout in the order your most likely to use them then it would help. I think its cross platform but I am using the windows version.

With this app you can load in a number of photos of a scene. You can add them one at a time or you can create a text file, list all the photos one per line and then save the text file and load in the list of images in one go. It would have been better to have used a multi select picker box but there it is - not a big deal.

The app comes with some sample images of a building and a text file list of them so load in that list. Once done it shows one of the photos and you can use the next and previous to scan through them.

The first job is to get it to match similar features between all the frames. On the menu there is "Matching" and then hit the "Start Matching" where it will bring up a progress bar and do some matching.

Once done it will cover the screen with a lot of little black squares. If you hover your mouse into these squares then it brings up a set of mini images showing where on the set it found that feature.

Next job is to get the camera calibration and positions from where the person took the photos. Under "Calibration" hit the "Auto Calibrate" item and wait for the next progress bar to complete.

Once done you can look at the scene in 3D by selecting the "Edit" menu item then "Mode" and "Overview mode" where it will spin the scene in 3D showing the camera positioned in that 3D space. Use the "Edit", "Mode", "Shot mode" to go back to looking at the photos.

Ok now it has enough info to find the coordinates. Hit the "Modelling" menu item and "Triangulate all virtices" item.

It should paint most of the little squares with a green dot. Actually this is a line passing though that coord in 3D space to all the features on all the photos. If they are in the same position on all the frames then its a dot. If it got it wrong then it will look like a line as it matches them in different positions. So in theory you should only use the ones with dots and lose the lines. Hopefully you wont get many with lines on them. It means your photos have too much movement in them or are too different.

Anyway, assuming that you have all the dots in green now, you can export them in 3D. Now it would be real nice if it could export in OBJ format but alas like a lot of apps it does not have it. In fact it only has a text file export. So hit the "File", "Export pointcloud (.txt" option and save the file somewhere.

We will need to convert that file into something more useable but ill show you how to do that later.

Anyway, you can see your vertices in 3D now by looking at it with a different view mode. If you go back to the "Edit", "Mode", "Overview mode" then not only do you see the cameras but all of the points in 3D and with luck they will form a collection which maps to a square looking side of building.

At this point the app devs would like you to select points in useful places one at a time and then make faces out of them. If you do that then you can build up your own models and it will texture them for you with superb texture maps based on the photos provided to give great buildings. Personally, I am more interested in generating 3D terrain so this is not an option. In the forthcoming posts ill show how you can make a 3D model from faces made out of that point cloud using Mesh Lab.
Tags: None Add / Edit Tags