Gradual selection???

[airmap3d]

Hello all,

I have asked this question before but sadly got no replies!  Maybe it's a tough one?  I'll try again...

Reprojection Error and Reconstruction Uncertainty... what exactly do they mean?  And what are the numbers on the slider bars representing?

I have basically worked out what I need to be doing with them to get good results but I don't actually understand the principle behind what I am doing!?!

ANY thoughts towards this topic would be very much appreciated!!

Thanks.

[RalfH]

+1!

Would be great if the Agisoft team could clarify this!

I also only have a relatively vague idea of what the numbers might mean, and being able to interpret them correctly would definitely help.

Also, it would be great if these parameters could be (optionally) added to the exported file when exporting dense point clouds.

[Alexey Pasumansky]

Reprojection error
High reprojection error usually indicates poor localization accuracy of the corresponding point
projections at the point matching step. It is also typical for false matches. Removing such points can
improve accuracy of the subsequent optimization step.
Reconstruction uncertainty
High reconstruction uncertainty is typical for points, reconstructed from nearby photos with small
baseline. Such points can noticably deviate from the object surface, introducing noise in the point
cloud. While removal of such points should not affect the accuracy of optimization, it may be useful
to remove them before building geometry in Point Cloud mode or for better visual appearence of the
point cloud.

http://www.agisoft.ru/forum/index.php?topic=738.0

[Porly]

Hello,

the reprojection error is the distance between the estimated and the real location of a point (in pixel). The error of the estimation depends on the quality of the alignment.

Example: You have got three photos aligned. One point was matched and estimated by only two cameras. Now it is possible to estimate the location of this new point in the third camera. Depending on the alignment-quality and the angle that was used for estimation, the point will not appear exactly on its true position.

So, in the alignment step the reprojection error is caused by alignment inaccuracy (estimated point) and the quality of feature extraction (matched/measured/marked point). Deleting these points will improve the optimizing step.

Reconstruction Uncertainty are TIE points (sparse point cloud) which are estimated in lower quality. This can be caused by point-estimation with a short base between 2 cameras or even a low overdetermination (point only matched in two cameras). Deleting these points has no affect to the optimizing step, since these points are valid matches (valid links between photos).

Hope this was the question

Best regards

EDIT: ups too late

[airmap3d]

Hi Alexey and Porly,

Thanks for the replies, very much appreciated.

It's all slowly starting to make sense.  Now to go and put it into practice...

Thanks.

[piste]

Hi everybody...
I reply in late at this topic.
As you can see in previous posts, here it's explained in detail the gradual selection workflow...
Now, with Photoscan 0.9.1, something is changed 
Gradual selection tool has been modified, I see two different options and one slide bar, but no more reprojection error and reconstruction uncertainty values...
So:

• How does it work Gradual selection now?
• At what step do I use it? Just after aligning images or after optimizing point cloud with GCP? (See aerial image tutorial in home page)

Thanks so much for every suggestion...

[Alexey Pasumansky]

Hello piste,

There are different criteria for point cloud and mesh view.
If you switch view mode back to Point Cloud you'll have reprojection error and reconstruction uncerainty options in the Gradual Selection list.

[piste]

Thanks Alexey!!
So, if I want to optimize my project at maximum I could follow this workflow?:

• Align images
• Create geometry as point cloud sparse
• Create marker and place them on right position in every image I can see them (filter images by marker)
• Import GCP coordinates
• Set proper coordinate system
• Optimize point cloud with marker enabled (images' positions disabled)
• Select point that are far away from object to reconstruct and delete them using crop tool
• Go to Edit > Gradual selection and drag the slider of error reprojection till about 1, then ri-optimize point cloud
• Repeat step 8 one or two times more till average reprojection error in "show info" in chunk properties is less than 1 or satisfactory.

Is it correct? 
Or is it better to delete points by gradual selection before optimizing point cloud with GCP coordinates?
Could I use also uncertainty reconstruction slider?
Thanks to everyone who shares your workflow...

David

[Alexey Pasumansky]

Hello David,

The workflow you've described looks ok. More or less similar way of optimizations were already discussed on the forum, for example: http://www.agisoft.ru/forum/index.php?topic=738.0

In case of large point cloud model distortions (similar to bowl-effect) it could be better to optimize the alignment using only camera coordinates and only after that use GCPs.

In your workflow I'd rather change points 6 and 7. And probably also combined new point 6 with the removal of points with huge reprojection error (there could be only few of them).
Also I do not recommend to remove too much points based on gradual selection criteria, since you can remove valid matches and the alignment will break after the optimization, causing geometry reconstruction problems.

Another quite important thing, based on our experience, is that average reprojection error higher than 0.8 pix indicates quite inaccurate solution. Whereas reprojection error less than 0.5 is almost perfect.

[piste]

Thanks Alexey for suggestions...
So my workflow changes in this way: step 1, 2, 3, 4, 5, (7+8+9), 6...

Another question about this:

In case of large point cloud model distortions (similar to bowl-effect) it could be better to optimize the alignment using only camera coordinates and only after that use GCPs

I'll never have any image coordinates available (we use camera without integrated GPS).
May I optimize point cluod only recognizing target on different images?... (image coordinates are important only in alignement process...)
Moreover, if I have no coordinates (nor images position or GCP known) because I don't want to georeferred my model, may I optimize point cloud only placing target in correct positions and then scale model?

Many thanks

David