building geometry is uncomplete

[jan]

The geometry which is build out of the 3d point cloud generated by the alignment process is uncomplete (see attachment). This is unpleasent because the computer had to run 5 days for the final unsatisfactory result.

What could be the problem? What can I do to avoid this. It is not the first time I encounter this problem. One loses a lot of time when processing lots of pics.

thx,
Jan   

[Monserrat]

Excellent work, I see it very well. 1440 Photos 5 days to is fine.

Thanks for sharing

[jan]

Yes I know, I am not complaining about the 5 days in itself, and the result which was calculated is also very good!!

However, it is a bit pity that only half of the area of the point cloud is used for the surface building here (the purple coloured area on the image).

I needed the whole area, because now I cannot calibrate the model: (I would place markers where no surface is generated from the point cloud), so I don't know the distance between two points, and I can not calculate this either. Therefore, this model is unfortunately useless for me.

The model was generated by a high end desktop for which I had exceptionally access for 1 week, but this week is over.

I was wondering why photoscan goes wrong here, because the point cloud seems dense enough to create a larger surface than it has created now?

kind regards,
Jan

[Monserrat]

Usually this indicates that a wrong reconstruction volume was selected. Too large reconstruction volume selections may be also undesirable as they result in larger processing time and memory consumption.

I understand your frustration, but this very clear in the manual, In some cases it is very hard or even impossible to generate 3D model of the whole object in one go. This could happen for instance if the total amount of photographs is too large to be processed. To overcome this PhotoScan offers the possibility of splitting the set of photos in several separate "chunks" within the project. The alignment of photos, building geometry and forming the texture atlas may be performed on each chunk separately and then the resulting 3D models may be combined together.
Working with chunks is no more difficult than use of PhotoScan following the general workflow. In fact,
in PhotoScan always exists at least one chunk and all the 3D model processing workflow operations are
applied to this chunk.

All you need to know to work with several chunks is how to create chunks and how to combine resulting
3D models from separate chunks into the one.

[ju523m]

Hi Jan,

I ran into the same problem just a couple of hours ago with a set of 390 images.  First I thought that there was an active volume selection, but there was none. I used heightfield / medium with a setting of 400000 Faces. Two opposite borders of the model looked  as if they had been cut. I was not able to reproduce the error with that data set. It was the first and up to now the only time that this error occurred.

kind regards
Thomas

[jan]

@Monserrat

Yeah I've read also the sentence:
"In some cases it is very hard or even impossible to generate 3D model of the whole object in one go"

I'm not sure, but this sentence might only apply when a computer is not "heavy" enough? For instance, with a low-end pc I've encountered out-of-memory problems, which is of course very normal.

However the desktop which I used for the model in the attach was a 24GB RAM; octocore with a GTX 480 graphics card, and it did never crash, nor had out of memory problems, so I'm not sure if the sentence above applies for this problem. Additionaly I've encountered similar problems in the past with less photographs, while at the same time having correct results with more photographs (on another desktop, and with smaller photosets).


Yeah I know about the chunks, it is the next thing I am going to try,

regards,
Jan

[gEEvEE]

Hi Jan,

I have a similar PC as the one you used (hexacore, 24 GB RAM, GeFore GTX 580) and I have this often. This has to do with the bouding box that is estimated by PS. What I normally do is a very fast reconstruction step (using fast, lowest) to see if a problem appears. If so, I choose to define the bounding box/volume manually. Afterwards, the problem is gone and the whole scene reconstructed.
Just out of curiosity: which settings did you use, as 5 days seems very fast for 1440 images. My PC could never do this using my reconstruction parameters and my GPU is even faster than your GTX 480. Just to know.

Cheers,

GV

[jan]

Hi gEEvEE,

Thanks alot for the usefull tip! I didn't know that one could define the bounding box/volume manually, I suppose it will be somewhere in the manual.

As for the software setting in Photoscan (alignment: accuracy high, preselect image pairs); (geometry building process: method, height field; quality, medium). For the other parameters  I just used default settings;

For the hardware setting, I used a PC of the vision lab of the physics department of the University, I don't know more details, but it was an octocore (or a quadcore in hyperthreading, also possible). I have no more access to this PC, but if you really want to know the hardware setting, you can send me an email, and I can give you the person's email responsible for this PC.

PS, I think we are from the same country 

[jan]

@GeeVee,

On the other hand, it might of course have been so fast (5 days), because it 'only' generated the surface for half of the point cloud

[gEEvEE]

Hi Jan,

yes, most likely we are from the same country (at work at Ghent University). OK, your settings explain the 5 days: I normally work on the "high" quality setting.
Cheers,
Geert

[Dmitry Semyonov]

Hello Jan,

We should recheck the automatic volume selection algorithm, but using a manual volume selection option should be a reasonable choice as was suggested earlier. Probably the problem was caused by a too small reconstruction volume selected by an automatic algorithm.

I would like to note that marker placement outside of the reconstructed surface is indeed possible. To do that you need to create a marker instance first, and then define its projections on all the photos where it is visible. More information is available in "To create a marker instance" and "To add a new marker projection" sections of the PhotoScan Professional User Manual (Creating reference points, Chapter 4).

Please let me know if you have more questions on this approach to marker placement.

With best regards,
Dmitry Semyonov
AgiSoft LLC