about the photo alignment

[NR1168]

Hi,
I want to take rice field images using drone. At the beginning of transplanting H85% abd W 60% overlapping worked. But Now after 2months it didn't work. So I increase the overlap ratio to H90% and W 70%. and this time it works for one field. But for  other field H90% and W75% still not working. So, what should I do now??

And for aligning the photos usually we use key point limit is 40000. Is there any problem If I increase this like 4000000. I found it can align photos more good??
Can anyone please explain about key point limit and tie point limit.

I need it urgent.

Thanks.

[Alexey Pasumansky]

Hello NR1168,

I can suggest to increase tie point limit rather than key point limit, for example, up to 10 000.
Also if you see that there are no tie points between the neighboring images you can try to Disable the pair preselection in the Align Photos dialog, if the dataset is not very big.

[NR1168]

Thank  you.

Can you tell me about the tie point and key point a little details.

Is there any specific overlap ration for higher vegetation period??

[NR1168]

And I want a good higher resolution orthophoto. So if I increase or decrease key point and tie point, is there any problem with resolution?
 

[Alexey Pasumansky]

Hello NR1168,

Key point limit
The number indicates upper limit of feature points on every image to be taken into account during
current processing stage. Using zero value allows PhotoScan to find as many key points as possible,
but it may result in a big number of less reliable points.
Tie point limit
The number indicates upper limit of matching points for every image. Using zero value doesn't apply
any tie point filtering.

So using very high number for key points may result in low confidence feature points used for matching that can affect the accuracy of the alignment results. Using very low limits for key and tie points may affect the stability of the alignment, as it would be performed on too few points, so both intrinsic and extrinsic parameters of the camera orientation may be estimated inaccurately.

The overlap should be sufficient to match the images taken from the neighboring camera positions irrespective to the flight direction. High and dense vegetation is usually problematic, since the greenery may look different form the opposite directions and almost no matching points can be found in this case.

[Mr Whippy]

Is there any way that Photoscan could juggle these variables iteratively and refine to an 'ideal' result over time?

I've always found it quite counter-intuitive that I have to find this ideal balance of variables myself.


I've increasingly tended to run with no Tie Point limit, and a fairly high Key Point limit.

I then duplicate the chunk, use gradual selection, and then Optimise Cameras, and so on, iteratively.

The only problem is you need to do a 'quick' dense reconstruction to check that there are enough Tie Points left for the dense reconstruction to work on all cameras... since there is no check (that I'm aware of) when Tie Point levels drop too low on a per-camera basis.


Couldn't all this be done automatically somehow? A way to get the very best camera position reconstructions and Tie Points, but without reducing them too much so dense point reconstructions can't work?
And/or warning that dense point reconstructions will not be accurate enough on specific camera image pairs?


As always, I find solutions, it just seems quite manually intensive and slow iterations when I feel that an automatic algorithm could achieve the same given some processing time?


Thanks

Dave

[stihl]

As a rule of thumb that I use (or at least that has worked for me for over 500 aerial projects, small and big) is that I only use about 20% of the initial amount of tie-points available.

The default parameters are 40.000 for the key point and 4000 for the tie point. I up these values to 160 and 120k respectively.
Then I use the gradual selection to remove bad points with pre-determined values. These values were determined by extensive testing with varied aerial projects which I prefer not to share publically for obvious reasons and because these values work with our setup and will probably yield different results with a different camera/lens configuration.

With this knowledge I find it easy to produce consistent Heightmodels and we've proven that with our work method we can reduce deviations for volumetric calcutions to an error margin of 0.5%

Don't go overboard with the gradual selection, as you've noticed yourself if you remove too many tie-points the cameras fall under a certain limit and will not be used for the dense cloud which leaves holes in your model.

I can suggest to take 1 data set and try the different alignment parameters and see for yourself what this means for the final product.
To test the heightmodels it's easiest to import them in a GIS and calculate the height differences between the models to see which one fits your needs the best.

[NR1168]

Thank you very much.
But still aligning is not working for me. I tried different key and tie point limits. also try to add marker and reposition and re align of selected images.

For high vegetation what are the other options to align the photos. I found usually it can be done by increasing overlap ratio. But for some case, by increasing overlap ration still it's not work. So, I want to know how I can accumulate images? Is there any other options?

[Vlad]

May be your photos is not good enough, low resolution that not allow find "features" required for tie point calculation?

Can you share 5-10 photo sequence for understand problem?

[Mr Whippy]

Thinking further from my previous post, I suppose the angle I'm coming from is a little like those for computer graphics rendering systems over the years.

Originally biased renderers were very popular because of very low CPU speeds, so you'd choose settings and the end result would be in line with the initial chosen settings...
...but with time unbiased renderers arrived which were very slow, but the longer you left them running the more accurate the rendering would become, the final render would never arrive, it'd just refine to infinity given time.

I think photogrammetry reconstructions could follow this path in time, simply left to their own devices for as long as desired and the results refine to ever finer details, rather than having arbitrary pre-set limits at the initial camera detection steps etc.


Hmmm

Dave

[Vlad]

I think photogrammetry reconstructions could follow this path in time, simply left to their own devices for as long as desired and the results refine to ever finer details, rather than having arbitrary pre-set limits at the initial camera detection steps etc.

What if I tell you that modern photogrammetry algorithms and tool work like you say. Unfortunately Photoscan still use old way.
If you interesting you can read this paper:
High Accuracy and Visibility-Consistent Dense Multiview Stereo HH. Vu et al. 2012.

I hope Agisoft one day stand from five point and do something to improve quality and speed of it not cheap app.

[NR1168]

Please check this link below to check my images.

https://www.dropbox.com/s/pli8oag318jblic/Images.zip?dl=0
The images were taken from a 11m height from ground. Using Sony alpha A5100, 24.3MP camera with 16mm focal length.