Agisoft Metashape

Agisoft Metashape => General => Topic started by: engelbeck on April 15, 2017, 07:56:56 PM

Title: Recommended Gradual Filter Settings
Post by: engelbeck on April 15, 2017, 07:56:56 PM
I don't know about the rest of you, but the documentation of Gradual Filter Settings does not clearly describe the meaning of the parameters, units, or recommended least a place for a newbe to start would be nice!  There are some hints elsewhere in this forum, but nothing direct and to the point.  What do you think about including a table like this:

Parameter                                  Units    Value                             Description
Reprojection error                    pixels     0.2       Statistical error in point placement
Reconstruction uncertainty     pixels      10       Possible variation in point placement
Image count                                  -            4         Maximum number of images having cross matching points
Projection Accuracy                 pixels      10       Accuracy of point placement from local neighbor points

Maybe as a community we can establish a place to start for the newbies (like me) out there.  Please review and correct as you see fit.

Title: Re: Recommended Gradual Filter Settings
Post by: Yoann Courtois on April 18, 2017, 05:44:49 PM
Hi engelback !

I'm currently trying to look through this tool and was also going to create a topic like yours.
I've several question about what you're presenting in your post:
- Are you sure about the unity of Reconstruction uncertainty and Projection accuracy parameters ?
- I'm not really sure about the meaning of the Projection accuracy parameter, may somebody know more about it ?

In practice, I'm actually using those following value while filtering my sparse clouds:
- Reprojection error between 0.5 and 1 pxl depending on the precision of the survey and the pixel size
- Reconstruction uncertainty between 50 and 200 in order to filter noise without loosing lots of points
- Image count not used: More than half of the tie points are match on only 2 pictures, thise filter would remove me to much data.
- Projection accuracy between 10 and 30 in order to filter noise without loosing lots of points.

Actually, I usually start with bigger values for those parameters, then adjust the model, and iterate while decrease values until nothing is moving so much.

Title: Re: Recommended Gradual Filter Settings
Post by: engelbeck on April 22, 2017, 10:47:56 PM
Hi Yoann,
   Thanks so much for taking an interest and responding. Honestly, I am not sure of anything.....I was hoping the Agisoft tech support Alexey Pasumansky would take an interest in contributing, especially correcting my post.

The definitions and values I posted here are my interpretation of the Agisoft Users manual and reading the following references. 

1.) Topic: clarification on gradual selection parameters please  (Read 12655 times)

2.)Photogrammetry tutorial 11: How to handle a project in Agisoft Photoscan
Recommends a reprojection uncertainty of 10
Reprojection Error of 1
Projection Accuracy of 7 to 10

3.)Best Lens for Photogrammetry
Routinely getting sub pixel accuracies between 0.4 and 0.6

4.)USGS Unmanned Aircraft Systems Data Post-Processing
Reconstruction Uncertainty Target 10, no greater than 50
Projection Accuracy 2 to 3
Reprojection Error Target 10% of the points selected
Tie Point Accuracy less than one, target 0.1

I was unable to find another article I saw where a professional photogrammetry company was achieving accuracies of 0.2 pixels with high end DSLR cameras and carefully selected lenses.

If you are only getting Image Counts of 2, I would suggest increasing the side to side overlap between images you are taking.  I make sure a feature on an object I take a picture of occurs in at least 3 images to get a 60% overlap...5 would give me 80% overlap.  When you add overlap from images shot from above an below you should see the Image Counts climb above 6.

Without any other direction I am resorting to experimenting with the values trying to improve the alignment and camera calibration to get qualitatively better results.
Title: Re: Recommended Gradual Filter Settings
Post by: Yoann Courtois on April 27, 2017, 10:45:20 AM
Hi engelbeck,

I've read your references with lots of interest.
However, I didn't find many deep explanations about Reconstruction uncertainty and Projection accuracy filters. I would say it's only some feed back for experiences :-[

So I would also like to have a bit of help by Alexey of those points :P

About image count: I'm aware about overlap factor in accurate reconsctruction. Indeed, I always use a minimum of 80% band-overlap and a minimum of 60% side-overlap. So I get quantity of tie points matched on more than 2-3 photos. Nevertheless, I never delete 2-photos-matched-tie-points using the filter, because it would reduce my tie point could size by two (or even more for more basic reconstructions), and this amout of points are really precious in order to calculate precise and global camera calibration, which is really important.
Title: Re: Recommended Gradual Filter Settings
Post by: Yoann Courtois on June 07, 2017, 06:23:18 PM
Up ! ;D