Stochastic Processes in PhotoScan?

[cbudin]

Hi,

Not sure if this question has been asked/answered before...

I'm using PhotoScan Standard with 8 cameras to capture human heads and bodies.  Generally the results are very good.

However, I have noticed that if I run Align Photos a second, third, etc. time I can get very different results.  Sometimes the alignment will be better, sometimes worse.  It would appear that there is some kind of randomness in the alignment process that in my case can have a very large effect on the resulting reconstruction.  It would be great to better understand where these stochastic processes exist in PhotoScan.

So, my questions are:

1) What stochastic (random) processes are at work in PhotoScan and where do they occur (ie which steps in the work flow)?

2) Is there any way to better control these random processes to get more consistent results?  I've already increased the Maximum Points per Photo from the default 40K to 100K, should I go even higher?

3) Which stages of the workflow might be improved by running them again, and which won't (ie don't involve random processes so should always produce the same result)?  I've already observed a fairly large variation in the Align Photos stage, which I assume has to do with choosing random points on the images as candidates for sample points.  Does Build Geometry also employ randomness at any point?

4) Does re-running any of the stages (such as Align Photos) result in refinement of the previous results, or does it always start over from scratch?


Thanks, in advance,
Clay
NYC

[Alexey Pasumansky]

Hello Clay,

The significant difference in the alignment results can be observed in case the camera positions and overlap produce some kind of instability. Also lack of feature points can lead to the similar behavior.

In case the object has good texture pattern and overlap is sufficient re-running Alignment stage will produce nearly the same results (here I can mention that running Align Photos stage doesn't use any information acquired on the previous run, it is Align Selected Cameras feature that is based on the previously found matches).