Modeling trees > Pseudo Points of Interest

[Patribus]

Hello,

I'm presently studying the possibility of 3D modeling real trees. The task is complex and difficult, but that is not why I'm writing.

The thing is: as I understand PS determines points of interest (POI) by analyzing the surroundings of the respective points.

When modeling trees a typical image you have are branches against a bright sky background. The scene is such that PS manages to identify POI in the white spaces between the branches (see image below).

Ironically this points seem to deliver correct results, since the image patch / structure where the points were identified move in a correct perspective from photo to photo.

My problem now is that these points appear in the end point cloud, as expected. But they to not mean anything in the context of the object.

So, what can I do against this? Creating a mask on the white regions is impossible. There are hundreds of this small regions per photo and over hundred photos. So I suppose that the workaround has to happen in the point cloud. These white (sky) point do also appear when generating the dense point cloud, although to some lower extend.

Some idea how to handle this?


Best  regards

[jtuhtan]

Hi,

I have had some success with modeling vegetation, and am also interested in developing a robust workflow for trees. The main problem you will have is also that the the tree branches are not completely still during your photography, which can also lead to a type of 'halo' effect around finer detail areas like leaves and small sticks.

https://www.youtube.com/watch?v=ystjSGqUBQI&list=UUNp0vouv7d04EqdXvH1phgw

One solution I have found is to use a stereo camera system which takes two images simultaneously, this helps to remove some of the spurious points.

You can use image processing software (I use MATLAB) to automatically detect and mask a decent part of the white regions, but a perfect mask is probably not feasible.

You can save the point cloud as a .ply file and then use a color filter to remove points within a specified range of RGB values. So far I have used such a filter to remove vegetation from point clouds and it seems to work pretty well...

Best,

Jeff

[Patribus]

Hello, thanks for your feedback.

The part of the matlab pre-treatment would interest me a lot. Is there some standard routine you're using or did you write some code?
Ironically the white patches between the branches and leaves seem to lead to correct results (assuming that there is no movement of the branches). I should check if results (from the point of view of the alignment) get better or worst when masking them...

Concerning the stereo cam issue: The thing is my case is, that I' collecting images using drones. So there is eventually the weight limitation. Which camera are you using?

An finally the thing with the cloud export. The problem about this is that it is not possible to re-import the data (edited point cloud) in PS. This issue is being discussed here: http://www.agisoft.ru/forum/index.php?topic=2466.msg14693#msg14693

Ahh, and finally, your video is nice, but let's say, I'm missing the tree there, hehe    I'm working with trees which go from 5 to 25 m and the method will be used (or shell be used) with trees up to 40m. You might imagine the challange this is....

Let's keep in contact...

Best regards