Looking for some tips in an unorthodox experiment - using web photos

[lemieszek]

Hi all,

I won't bore you all with details of why I'm doing this, but I'm trying to reconstruct the Notre-Dame de Paris with nothing but photographs downloaded from Wikimedia Commons. This means all kinds of cameras, settings and positions with photographs that are literally decades apart, with and without the roof (RIP).

So far I've been pretty successful, all things considered. I've matched 1.1k images (out of 2.8k I downloaded). But I'd like to see if anyone has suggestions to make the end product more accurate.

One bump in the road I've hit is that some photographs just won't match correctly. The attached image isn't showing this well, but the sparse cloud shows some "ghosting" of the west (main) façade - there's an alternate version of it that just continues straight as if it it were part of the south façade (Imagine you're looking at a plan - where there should be a L, there's a _L, where the _ is a ghost west façade aligned south.

I've created 14 markers and marked each of them on multiple photographs - as many as I can in each one. I tried finding all pictures that show this corner of the building so that markers on both façades are mapped multiple times. I'm not sure though what the best practice is after that - do I just optimize cameras, do I run "gradual selection" and take out some bogus points beforehand, or do I reset alignment and run it again? I've tried a little bit of each but no luck so far. The sparse cloud seems to be getting worse, not better D=

Any tips or questions are welcome.

Cheers.

[SimonBrown]

If I recall, a similar approach was used to model the Treasury building in Petra, Jordan?

First question; what do you mean by "accuracy"?

The scaling/measurements in the scene?

The geo location?

Or the image alignment?

[lemieszek]

If I recall, a similar approach was used to model the Treasury building in Petra, Jordan?

First question; what do you mean by "accuracy"?

The scaling/measurements in the scene?

The geo location?

Or the image alignment?

I didn't know they'd done that in Petra. I'll look it up!

Ideally I'll eventually have it reconstructed with accurate scale and position, but at this time my challenge is alignment. Take a look at the attached sparse cloud - you can clearly see ghosting, plus Metashape still "thinks" the west façade is aligned with the south façade. Even the positions of the markers in the sparse cloud are iffy, even though, see other three images, I feel like I've given it enough information to understand the geometry of the cathedral.

I just tried resetting alignment and redoing it with the markers in place - it took it all night and got a much worse result.

Some of the photographs are geotagged - maybe that's screwing up the result? (I haven't disabled them as reference in the hopes that they'd yield fairly accurate scale and position...)

[olihar]

Look back at old Photosynth.

[SimonBrown]

The absolute and best path to accurate photogrammetry is to make sure the source images are shot with the end game in mind...are technically correct...and carry the necessary overlap. This priniple is hammered home in our https://accupixel.co.uk/courses/metashape-foundation-course-for-forensics-surveying/ course.

But in this instance it's not the case. First suggestion would be to review the images and select a set that appear to be shot from the same distance-to-subject. Then see if there is any correlation or grouping of camera and if there is a common type you could try putting them in a single sub-folder, or folders if there are a few common groupings.

Forget using any embedded GPS at alignment. Camera phone accuracy causes more havoc with alignment so turn it off. Maybe select some after alignment and recursive optimisation.

I would hesitate to use markers to align. This is a bit advanced (we teach it in the Pro course) and its best used when there is some common overlap but not enough tie points, such as when working in areas of vegetation.

Clean, uniform source images will always work best. But if not available then we work with what we have...so give it a go.

[Alexey Pasumansky]

If I recall, a similar approach was used to model the Treasury building in Petra, Jordan?

Also something similar to Project Mosul (now known as Rekrei):
https://rekrei.org/gallery