aligning chunks by point-based method and effect of dense cloud quality

[octopus]

Hi. I am trying to align multiple chunks (Workflow>Align Chunks command) by using point-based option and have a question.

Does the point-based method use tie points, or does it use dense cloud points of source chunks?
    Currently, my chunks are processed with high-quality photo-alignment option and low-quality dense-cloud generation option.
    To get a good chunk alignment, is it better to re-process each chunk with high quality dense cloud generation, or
    it does not matter if the aligning calculation relies on tie points only?

Also, in either way, does it help to clean-up by manually deleting noise data (inaccurately placed tie points or
   dense-cloud points generated around the edges of the model) before running the chunk align command?

I would appreciate any advice very much

[Alexey Pasumansky]

Hello octopus,

When you run Align Chunks operation using point based method, PhotoScan will try to find the matching points between the images from different chunks. In case you have key points kept in the project they will be re-used (but re-matched), otherwise the procedure will start from the feature points detection.

The presence or absence of the dense cloud in the chunks has no effect on the chunks alignment.

[octopus]

Hi Alex;
Many thanks for your explanation. Now I have a better understanding on chunk alignment.
I managed to align my chunks satisfactly with point-based method.

If I may add one further question, does the two methods below make any difference?
Assume there are 4 sets of photos, set A, set B,  set C and set D, each with several hundreds photos. And

 method 1: Import photos of A, B, C, D into one chunk and align all the photos together in one operation.
 method 2: First, for each of A, B, C, D, align photos and make a small chunk. Then align these
                   4 chunks (A, B, C. D) and merge them together to make a chunk including all the cameras.

I imagine method 1 would adjust each camera position considering all other photos, while method 2 positions
each camera only relative to the photos in each set. Therefore

 a.  method 1 takes significantly longer time for photo alignment
 b.  method 1 camera positions are much more precise than method 2.
 c.  method 1 also requires much more RAM for photo alignment

Are these points correct?

My computer has limited amount of RAM, and has difficulty in method 1. I am hoping method 2 makes OK result.

T

[Alexey Pasumansky]

Hello octopus,

In general your assumptions are correct. Method 2 is less accurate due to the fact that the chunks (during chunk alignment) are treated as a solid models to which only Affine transformations are applied - rotation, translation and scale, so separate alignment of the image subsets may result in some artifacts in the areas where the chunks are overlapping.

As for the RAM needed, you can reduce the memory consumption by enabling the Fine-Level Task distribution in the Network Preferences tab (currently available only in the Pro version only).

[Mak11]

Alexey,

Does the Fine-Level Task Distribution function means that one can, for exemple, build a dense could, even if on a project that was running out of memory before hand (on a single PC)

Regards

Mak

[arose]

Is method 1 still the better option if the photos chunks A, B, C, and D are each photos of the same site but different years?
In this case, it seems like method 2 would be better.