How to optimize Reconstruction
I'm working with some imagery that introduces some weak spots in a Reconstruction. Cameras align, but sometimes only with help of Control Points, which already is a warning signal to insufficient number of photos or with inadequate overlap. I'm scraping by in a couple areas, meaning the cameras align with help of Control Points, but I see the distortions in the model, slight bowing in a wall as well as a slight separation in a wall. So, I've got a few ideas about how to optimize for such issues, thanks for checking my logic against how RC in fact thinks.
I see a setting for changing the Weight of a Control Point, from the default of 10. to I suppose 100.0 being the highest? If I'm ultra careful about setting Control Points and highly confident about common features, then what's the harm in acting on that confidence by setting Weight to 100?
If you're highly confident of strong Control Points, but see high reprojection errors, but also knowing this is in an area with distortions, how best to relax the Component to then really use these Control Points to correct the distortions, thus bringing down the reprojection errors to acceptable values?
If Merge Components Only = False and Force Component Rematch = False, then I'm unclear what that means for established features. That is, say I've aligned 2000 images, 50 remain unaligned from the previous Alignment, I add Control Points and now have 2025 aligned. With the stated Alignment Settings and now adding more Control Points, will RC remain flexible in readjusting camera poses as it tries to integrate these remaining 25 images? Would clearing cache have any bearing on the new Reconstruction? This is important because if I already see distortions/artifacts developing, these remaining photos (with the help of control points and possibly with high Weight settings) might help RC to better triangulate the surrounding geometry, but I'd that would only work given plasticity in the Component. I also wonder if the order in which I introduce problematic imagery might help RC to not "bake in" distortions or drift by not throwing too much at it at once or robbing it of a benefit in using less problematic imagery first to help form a better triangulated model before it then has to deal with the problematic images.
I hope this makes sense, and thanks for your insights.
Benjy
I see a setting for changing the Weight of a Control Point, from the default of 10. to I suppose 100.0 being the highest? If I'm ultra careful about setting Control Points and highly confident about common features, then what's the harm in acting on that confidence by setting Weight to 100?
If you're highly confident of strong Control Points, but see high reprojection errors, but also knowing this is in an area with distortions, how best to relax the Component to then really use these Control Points to correct the distortions, thus bringing down the reprojection errors to acceptable values?
If Merge Components Only = False and Force Component Rematch = False, then I'm unclear what that means for established features. That is, say I've aligned 2000 images, 50 remain unaligned from the previous Alignment, I add Control Points and now have 2025 aligned. With the stated Alignment Settings and now adding more Control Points, will RC remain flexible in readjusting camera poses as it tries to integrate these remaining 25 images? Would clearing cache have any bearing on the new Reconstruction? This is important because if I already see distortions/artifacts developing, these remaining photos (with the help of control points and possibly with high Weight settings) might help RC to better triangulate the surrounding geometry, but I'd that would only work given plasticity in the Component. I also wonder if the order in which I introduce problematic imagery might help RC to not "bake in" distortions or drift by not throwing too much at it at once or robbing it of a benefit in using less problematic imagery first to help form a better triangulated model before it then has to deal with the problematic images.
I hope this makes sense, and thanks for your insights.
Benjy
-
Do. It remember exactly what's going. But if I remember right. RC will clean all calculated information and will use pose/lens/etc priors (if the exist).
So result similar to like create new project import images and run alignment. Except it will reuse detected features from cache if you did not change some of advanced settings. -
Ah. Phew!
Somebody should write that down in a neat little diagram... ;)
Thank you again!!!
Here's the result of the first try as you suggested with my roofspace.
Almost perfect, even the rafters almost line up at the loose ends, thats like 2-3 cm on 8 m long arms in Z direction!
Before that, I did have quite a big issue on the lowest storey.
No idea why I didn't like it at the time I first tried it...-
1304Grouping result.jpghttps://www.capturingreality.com/forum/download/file.php?id=13042097281imageFile jpgGrouping result.jpg
-
Please sign in to leave a comment.
Comments
34 comments