clarification on gradual selection parameters please

[EvanThoms]

Hi there, I am looking for some guidance on the best use of the gradual selection properties to select and delete noisy points after the 'align photos' process with version 0.9.0.

Say, for example, that after I align photos, the chunk info shows the reprojection error to be 0.379438 with a max of 1.1552. What are the units on these values? Are these RMS errors? Is the first number the average?

Given those numbers, what's a statistically reasonable number to choose for selecting points in the gradual section dialog based on reprojection error? When the window pops up, the value in the level field is 0.681566. How was that value calculated?

Likewise, what are the units on the reconstruction uncertainty?

Sure, I can move those sliders around until I get a selection that 'looks right', but I am dealing with uneven terrain. I don't really know which points have been correctly located and which have not. It introduces user bias. I would be happier if there was a way to select, for example, only points that fall within a 95% confidence interval.

Thanks

[FoodMan]

glad you asked, I'd like to know myself too

[Alexey Pasumansky]

Hello Evan,

"Reconstruction uncertainty". In case you have only two cameras and a point is being triangulated by intersection of two rays there is a direction in which the variation for the point position is maximal and another direction with the minimal variation. Dividing one on another (max to min) we've got reconstruction uncertainty value. Mostly this criterion is designed for visualization and estimating the errors. In general this value characterizes the accuracy of positioning points in cloud.

"Reprojection error" demonstrates the accuracy of point positioning and is specified in pixels. We recommend to remove points with huge reprojection errors before optimizing photo alignment.

[piste]

Hi Alexey...
I have yet some problems in understanstanding how gradual selection works...
After photo alignment which points have to be delated or considered wrong?
How could I set Reprojection error and reconstruction uncertainty?
If you could tell me what would be the corect workflow step by step I would be so happy 
Thanks

[FoodMan]

"Reprojection error" demonstrates the accuracy of point positioning and is specified in pixels. We recommend to remove points with huge reprojection errors before optimizing photo alignment.

oki... mmm, am I missing something...? you remove (delete.?) some points..... and then do an optimization...? how do you do it..? I mean if i do an "align photos", Pscan will start to re-detect the points...

thank you.

ps, btw, can you also remove uncertainty points..?

f/

[FoodMan]

Well if we can't optimize in the base version..... what is the point to use the gradual selection ...on points...?

f/

[Alexey Pasumansky]

Hello piste,

Before using optimization tool we recommend to remove obviously incorrect (e.g. far away) points from sparse point cloud and to check the reprojection error, removing points with great reprojection errors. However, removing too many points is not recommended since you can loose correct matches and break up the alignment.


Hello FoodMan,

Optimization of the photo alignment is currently a feature of the Professional edition only.
Removing points (and therefore valid matches) from sparse point cloud may help in depth maps reconstruction if there too many outliers or incorrect matches. Also it could speed up the processing since depth maps are calculated only for images with the sufficient overlap based on the matches.

[piste]

Hi Alexey...
maybe I haven't been so clear...
I want to know: how set the reprojection error value to clear wrong point,
and the same thing for reconstruction uncertainty...
I read that reconstruction uncertainty value is given by Max/min error in direction of point positioning...where could I find these values?
Sorry but I continue to not understand the statistic theory on which are based these two filters...
I don't want to filter point after alignment only watching them while they are coloring in brown... 
Thank you very much

[Dino]

You can find those values under each chunk: Right click - show details...

[piste]

ok, thanks dino...
so, if I have understood, I have to see what is max/min value and then set this value as reconstruction uncertainty filter?
And for reprojection error? Is there any statistical value to consider?

[gEEvEE]

Hi,

This is more or less a good workflow.

1. Align all your images. Once this SfM step is finished, you will end up with a sparse point cloud. The reprojection error for that chunck gives you the average and maximum value of all reprojection errors. Ideally, they should be lower than 1 pixel. So, now one can go and edit the point cloud;

2. Select points that are far away from the object you want to reconstruct. Now delete those points (better: select the object and some space around and delete - using the crop tool - all other points).

3. Go to Edit > Gradual selection and check how many points are selected when you drag the slider till about 1. If there are just a few hundreds till a few thousands of points, go ahead. Now the points with a reprojection error higher than 1 pixel are selected. Delete these points;

4. Now that you have a sparse point cloud containing only points with a reprojection error smaller than 1 pixel, optimize the point cloud (which runs a bundle block adjustment and also optimizes the camera parameters). When doing so, uncheck "Fit Aspect" and "Fit skew", since these are 0 in normal frame cameras.

Now, go to your chunck, right click and select "Show Info" to verify that the average reprojection error dropped. Most likely, there will again be points with an error higher than 1 (since all points and their reprojections are recalculated during the optimization step). You could start again from step 3 and make 1 to 2 more iterations until you are satisfied with the global accuracy.

As was already mentioned: never delete too many points, as image alignments can be broken. Never iterate too much: the accuracy values will become better and better, but your reconstruction will show ripples. Best is to optimize 1-3 times and try to get as little points as possible with a reprojection error surpassing 1.

Hope this helps. CU,

Geert

[FoodMan]

ahh ok, that was what I thought ... thanks Alexey.

f/

[piste]

Thank so much gEEvEE!!
You're great 
This is what I wanted to listen...
Bye
David

[Matt]

Hi Geert,

I have been more or less following your workflow as prescribed in an earlier post from April this year.

"My preferred workflow is to align everything using 20 000 points and afterwards filter the points based on reprojection error. After deleting the points with a reprojection error higher than 0.5, I optimize the point cloud and filter again, until the end results delivers me a maximum reprojection error of 0.5."

Do you manually trim the obvious outlier points first or just leave it to the reduced bounding box and  reprojection error to select them? At present i delete them first.
Do you still aim for a reprojection error of around 0.5 where possible or have you settled on a maximum reprojection error of 1 as to be suitable for high accuracy work?

Thanks for all the great papers. 

[DCK]

Can someone explain to me the scales used for reprojection error and reconstruction uncertainty. For instance, what is a reprojection error of 1.2? What is reproduction uncertainty of 0.5?

Thanks.