In about 30% of our projects (turntable/lightbox), alignment in RC fails, and our current solution is to run the alignment in another software (almost never fails), then load the bundled.out file in RC to complete the process. I wonder whether there is a set of RC alignment settings that is closer to the this software’s process. Even if it is very slow (like in that software) it would still be really nice to have this option in RC.
Hi Adam,
have you tried the general advice in the FAQ? Do you use exif grouping?
Try increasing ‘preselector features’; by default RC keeps this value modest to make it fast. This value determines how many points will be sent to matching: unless your points are quite noisy, having more of them has a direct, positive relationship to correctly relating views.
Increasing ‘preselector features’ will scale RAM usage with it, and in some cases (for me, at least), values over 20,000 will lead to core dumps for large scenes (3,000+ images).
I’d say the ‘max features per image’ should probably be a multiple of 2-4 of your preselector setting. The latter variable sets how many features are generated, the former sets a ceiling for the number of those found to be passed to the matcher.
Hey Kevin,
that’s some really good explanation - has this been added to the help by now? Or where do you know that from?
Because to be honest to me "pre"selector means that something is only temporary, for a first step. Never would I have guessed that it is the actual maximum number of used features.
It could also be very valuable for the discussions of the crashes we are discussing elswhere.
A similar way to go would be to raise the Max Reprojection Error, which could also increase the number of used features.
I never used the other software so I don’t know for certain but I could imagine that they group images by focal length automatically or that all cameras have to be identical anyway. This is why I brought up the grouping - it can improve the initial results a lot.
Like you, I took the pre in ‘preselector’ at face value until I noted some of Wishgranter’s posts from a couple years back, notably:
‘Preselector features - how much from the max features will be used in alignment’
I’ve found that Wishgranter’s assessment is correct, increasing the ‘preselector’ value changes the feature count output to the matcher while changing ‘max features per image’ does not.
I agree completely that the wording would lead you to conclude the opposite!
Thank you so much for that pointer! Incredibly important!
Somewhat embarrasing that I participated as well. At least it is long enough ago to account for it to some degree… 
Thanks all. Preselector was the key. I increased the preselector value to 40k (=max features per image) and now projects that had previously failed are aligning as expected. This is great, as Reality Capture is much faster than the other alignment software were using for those cases.
Glad to hear it Adam!
It would be interesting to compare: keep your preselector value at 40k but increase total points to double that, to see if the matching gets passed higher quality points and therefore yields a better reconstruction.
If not, then there’s no reason to find more features than will be used, and we should all keep preselector and total points per image in lock step…
I always thought I read between the lines that feature detection already works on a quality basis. On the other hand, the software should not be able to predict how efficient features on a 2D image will turn out as 3D tie points. So it could be worth a try…