marker vs. point based chunk alignment

[toddb]

Does anyone have recommendations about which alignment method is more accurate or seems to work best for large projects with hundreds of images?  I currently have about 1000 images divided into eight chunks.  I've placed markers throughout the images to set the coordinate system and also used these in a test chunk alignment.  The alignment using the markers does not seem to position the markers in the same location (within the 3d view of the point cloud).   I'll try to include an image that shows what I'm seeing.  The markers are set in a 10 meter grid throughout the landscape but the 3d view shows markers with the same positions several meters apart.  Will this difference in position carry through to the model and orthos or is this somehow adjusted when the geometry is built (my guess is that this will carry through but I wanted to ask)?  Would it be better to use a point based alignment instead of the markers?  I've read on the forum that there are some issues with large projects and point based alignment.  Should I download and install the pre-release as I've also read that there are some adjustments that are being made in that package?  If you have any suggestions I would very much appreciate it.  I would like to build the geometry but want to make sure that the alignment is set as this process will probably take several days on my computer.
Thanks,
Todd

[Alexey Pasumansky]

Hello Todd,

Are there any reasons why do you prefer to split the set into chunks (for example RAM limit)?


As for the alignment actually there are three ways for the chunk alignment - apart from the marker and point based chuck alignment, chunk would be aligned automatically if they are georeferenced. If all the chunks are in the one geographic system you can use "Show aligned chunks" button on the Toolbar to see their relative positions in space.
Sometimes it is better to align chunks using their geographic positions than using other methods.

The difference in the marker positions (when using marker based chunk alignment) could be caused by the difference in camera alignment in different chunks. While the chunks are aligned only afine transformations are used, so maybe it is better to perform camera alignment optimization (if there are enough precise GCPs on each chunk) - it could help to correct deformations.

[toddb]

Hello Alexey,
I only have 8 gigs of RAM on my computer so I'm somewhat limited in terms of memory for processing the images.  All of the chunks are in the same geographic system with 20+ GCPs (usually more than 30) per chunk.  I had not tried the show aligned chunk option.  Does this use the same method for alignment as the align using markers?  The result looks similar but I'll check against an earlier version of my file just in case my marker alignment is causing this similarity.  Where is the camera alignment optimization?  I couldn't find this in the help.  If this will fix the offset that would be great.  Thank you for the information.  In general then is it best to either use the marker alignment or show aligned chunk option when dealing with geographic data with plenty of GCPs?
Thanks again,
Todd

[Alexey Pasumansky]

Hello Todd,

According to our calculations 8 GB should be enough for 1000 standard aerial images processing.

The difference between marker based alignment and referencing chunks using Ground Control is the following: when chunk is being aligned to the reference chunk (in case of marker based alignment) estimated marker coordinates of the referenced chunk are used, while during georeferening source data is used and it is more accurate.
In such cases with several chunks and lot of GCPs we recommend to georeference and optimize each chunk separately instead of aligning them using markers. The later way could give additional accumulated errors and is not so accurate.

To optimize camera positions you should use Optimize feature in Ground Control dialog/pane (corresponding to the PhotoScan version used). Please refer to the Optimizing Point Cloud section in PhotoScan manual (pp. 26-27).

[toddb]

Hi Alexey,
Thanks for the suggestions. I decided to try to process all 1000 images as you noted that it should be possible with 8 gig ram. The processing time however is currently listed as 360 hours remaining after 8 hours of processing. Does this seem correct and if so would preselection of the images significantly decrease the time? I have an i7 processor with 8 gigs of ram. I doubt that my video card will be of much help. It's not the best system for processing but it is what I currently have available. Given the time would you suggest proceeding with the full 1000 images to improve alignment or would I be better off to optimize each chunk and then try to improve alignment across these chunks?
Thanks again for your help.

[Alexey Pasumansky]

Hello Todd,

Yes, using preselection will significantly reduce processing time. When processing large aerial datasets it is recommended to use at least Generic preselection or Ground Control if there is GPS data for cameras - it is better and should be faster (in case of Ground Control preselection correct coordinate system should be defined in Ground Control pane settings).


If there are enough resources splitting into chunks is recommended only in case when all images could not be aligned (for example, when there is wide river dividing the area) - than each chunk should be processed separately.

So with preselection this project should be aligned in several hours 6-10, depending on the CPU (since GPU is not involved in photo alignment stage).