decrasing the number of tie points : how is it done ?

[Bruno Andrieu]

Hello
In the 1.1 version, there is now a tool to reduce the size of the sparce point cloud
("demarrier les points unitives" in the french version).

There was already the  "gradual slection"  to suppress outliers.  But I do not see how the new tool proceed.
can you explan the rule used to chose the tie points that are eliminated by this tool and in which case is it useful ?

Bruno

[Alexey Pasumansky]

Hello Bruno,

This new tool reduces the number of tie points per photo up to the defined limit by deleting most unrelieable matches.

[Bruno Andrieu]

Thanks Alexey


And does it make a difference wether we limit the number of tie points during the alignment process ( the new option in v1.1)  or we thin the point cloud a posteriori  ?

[Alexey Pasumansky]

Hello Bruno,

Using Tie-points limit option in the Align Photos dialog has identical meaning. But if you use this option during processing, some time may be saved during estimation scene structure step.

[Bruno Andrieu]

Thanks  for the explanation

Does photoscan compute some value of likelihood when  associating  points from two different images ?

It would be nice to have  a  global value for the sparse point cloud that would reflect the " average" likelihood of the tie points, or even to give some statistical  distribution.

I try to have the "best" sparse points clouds to optimise alignment. So  I proceed with removing outliers, selecting points identified on more than 2 cameras, and now there is this new thining option.  I would find useful to hace a measure of the achieved quality

It would be great to have a measure of  accuracy of the sparse point cloud  and  even (great +++)  of the information brought for alignment ( to choose between less but highly accurate points or more points with lower accuracy ? )

Bruno

[jo]

Hi bruno,

Exactly like you, I try to have the best sparse points clouds to optimise camera alignment. And I am so very interested in understanding the functionning of the great new feature of photoscan 1.1 regarding the selection of "best" tie points. But I think there is currently no agreement on what would be the best set of tie point for optimization, apart from the fact that it is better to have a limited number of tie point well distributed on every images and viewed on more than two cameras than having a large number of tie points seen only on two cameras and  localised on a specific location of the images.

Thanks, to Alexey Pasumansky's help,  I exported tie points of photoscan 1.1 (generated with a tie point limit of 1000 per camera) into another software (micmac from ign France) in order to compute the residual for each tie points. What I have found out is that tie points residual, which is often used as a quality measurement, is less good than for tie points generated in photoscan 1.0.4. On the other hand, about 80 percent of  tie points computed with photoscan 1.1 were shared by more than 2 cameras.

In summary, I think that the  "gradual slection"  to suppress outliers in the sparce point cloud is (was) a bad approach, because tie point residual is not a fair measurement of its quality. The quality of cameras alignement should be checked with external measures (ground control points or embedded GPS/camera location).

Jo