Need help w/ alignment and merging chunks...

[Integr8d]

Hi all. This is a little urgent. I only have 2 days left in my evaluation! I have been reading the manual and trying to troubleshoot with the already-created forum posts. But I don't seem to be getting anywhere.

First, please observe this link to my project...

https://mega.co.nz/#!Q9BU0Q6b!rRO3YCUSL92td53dsbWpM3EnMrbSFPytnLUk0XD4yos

If anyone has the time/resources to help, I would greatly appreciate it.

Now, what I have in that dataset are photos of a movie theater that I was recently using for a screening. I wanted to model it. And so I took my Canon Rebel 550D and tripod to the theater. It was relatively dark. So I had to use long exposures and a shutter delay to remove vibration. My kit lens was set to full wide (18mm), manual focus, VR: 'off' and aperture priority (either f8 or 11). I let the camera pick exposure. The results came out okay. I think the ISO was set a little higher than I would've liked. But all in all, I could live with it.

Chunk 1 has the pics from the front of the theater (closest to screen, pointing to the rear). Chunk 2 has the reverse of that (rear of theater, pointing to the screen). There are probably another 200 photos taken from the sides and from other vantage points (which were excluded due to size)... I am able to align the images in the chunks to what I would assume is a relatively high degree of accuracy (thanks marcel http://www.agisoft.com/forum/index.php?topic=3559.0). I am able to render out the point clouds and am happy with the results. But where I seem to be needing assistance is with the alignment and merging of chunks. I assume that aligning the chunks is a prerequisite before merging. So I align them w/ a Point Based method, high accuracy and the default 40,000 point limit. Chunk 1 is referenced... I merge the chunks based off of the dense point clouds. The process happens very quickly and I'm left w/ a merged chunk. It contains all 55 cameras, w/ 52 of them aligned (I disabled 3 of them). There are tie points and a dense cloud. However, the points and cloud are only from Chunk 1. And I believe that the cameras are not aligned as a whole. But rather as 2 sets.

My earlier attempts included using markers for both chunks. As a side note: the ray plotting is very accurate, out of the box. And it only seems to get better, as you refine the positions... I used the same marker names for the corresponding points on the 'front' photos and 'rear' photos. However, that didn't seem to help very much w/ alignment. As far as I can tell, 'aligning' chunks isn't doing anything other than place an [R] next to a chunk. Ultimately, I know that I'm doing something wrong... I'm also hesitant to use markers, as they're not available on the standard version.

So at this point, I'm at a loss. You can see, in my project, that I've taken it to the point where I've created both chunks. And on their own, they seem fine. What I would finally like to do is figure out how to align most, if not all, of the cameras and then to create a DSM from them. It's a tricky project because the photos from the sides of the theater look almost exactly alike. And when I though that Pscan would just magically create this model for me, I fed it all 220 photographs. And what came out of the other end, 13 hours later, would not be considered anything more than abstract computer art

If anyone w/ the standard version could especially chime in, I would appreciate it. For a hobbyist, the cost of the professional version forces me to look at other options (Pmodeler Scanner being the next area of focus). It would be nice, if there was a $1500-2k 'prosumer' feature set.

Lastly, I'm running this on a custom-built PC w/ the specs: i5 4670K, 8GB memory, 2x AMD R9 290 GPUs. Outside of the memory limitation, the software performs well. So upgrading the memory is not a problem. I'm looking at replacing the CPU, motherboard and memory though. Anandtech's Haswell-e review leads me to think that the mid-range chip might be a sweet spot for price/performance. If anyone thinks otherwise, please let me know.

Thank you for reading.

[Alexey Pasumansky]

Hello Integr8d,

You can merge only aligned chunks (they will be marked with [R] sign in the Workspace pane), so if after the chunk alignment operation you have only one [R]efenced chunk, you would not be able to merge the chunks as expected.

Point Based chunk alignment actually starts the image matching procedure between the different chunks, at it's quite time consuming operation. So probably Marker Based approach would be better (you can use a Trial key for Pro to save the proejct if it works). But this method requires markers with the same labels to be placed in the different chunks, but pointing to the same real world feature.

[Integr8d]

Thank you, Alexey. I tried that and it works Interestingly, the problem now must be a scale issue. I created a third chunk and aligned a set of photos that travel along the 'stage right' side of the theater. Surprisingly, Pscan was able to line them all up! But there seems to be a new issue...

To summarize the question: What did I do wrong, when I mark corresponding real-world points, in separate chunks, I then align and merge the chunks, but the output has two of the same markers in slightly different positions? Time to close program and restart?

[Alexey Pasumansky]

Hello Integr8d,

Markers with the same labels and very close positions indicate that PhotoScan was unable to fit chunks perfectly (due to slighly different model distortions, for example) and thet you haven't used Merge Markers option for chunk merging operations. If the option is used, markers with identical labels will be treated as the same marker and the projections from the different chunks will be merged.

In principle, you can try to optimize the camera alignment in the merged chunk then, as it could improve the result, however, optimization will discard dense cloud and mesh model, that might be probably unacceptable.

[Integr8d]

Yeah. I don't know if it's buggy or what. I've made a lot of progress today. But the software just started acting weird. I have 5 chunks (4 of which I've aligned successfully). The 5th one, when I try to align it with the others, and please keep in mind that I've given it at least 5 markers in commonality with another chunk, will not align (give me the [R] symbol). When I try to align, chunks 2-5 (I was assuming there was a max of 4 chunks), the [R] disappears from 2 and 3. And now only 1 and 4 are aligned, according to the software. Are there limitations to the number of chunks? With all 5 chunks open, the program only consumes 150MB of memory. So that can't be the limiting factor.

[Integr8d]

Alexey, in addition. Something just started happening w/ the view of first chunk. I could see the sphere, and the camera panes and marker dots. They were all tightly bound together and when I tried to zoom in, the just flew by the 3D window. I zoomed back out and tried to reset the view. But that was all it would show me. I rebooted the computer, thinking it was a memory error. No dice. I'm now realigning the photos.

[Alexey Pasumansky]

Hello Integr8d,

Unstable results of chunk alignment operation means that there are problems in finding the matching points between the images from different chunks. So there's no max number of chunks PhotoScan can align at once.

As for the display issues, it might be caused by some camera outlayers that are not properly aligned. I can suggest to use Look Through option for any camera.