Regular followers of the Vobster Photogrammetry Project will know we have been working on this for some time. The project is all done in our spare time, for free and for the sheer joy of pushing the boundaries of photogrammetry.
The game changer has been the introduction of UWIS and underwater GPS. To keep the project running efficiently we needed two things; scale and navigation. The success of scale and geolocation have been discussed here and here respectively…and underwater photogrammetry is never going to be the same again…
But what about the masses of data we already have? None of the existing underwater images are geolocated, so does that mean they are junk and it’s a reshoot?
Lazy? Or Efficient?
We do have a lot of data. We even ran a “guess the number of aligned images” a while back with the model hitting just shy of 38k images aligned. So how could we reuse what we have?
One approach is to only scan the quarry floor and base of the attractions. None of the source images of the existing models will be carrying GPS data or indeed scale, but if the quarry floor images do and act as master when we align chunks then everything should work.
Which is exactly what we set out to prove recently. The targets were the wheelhouse, helicopter, caravan and the yacht Poppy – all close enough to to swim over a couple of short dives and all have been separately modelled in the past.
Why not do one long dive? Well yes, we could. But until we prove concept the work there is little point generating piles of data only to find its junk…
Overlap Compromised
The section between the wheelhouse and helicopter aligned perfectly. We also included a small section of Jaquin bow for later reference, but the first dive was a hit.
The second dive was less productive. A few gaps in the parallel runs meant full alignment didn’t happen. This is typical when distractions occur such as “Where is my buddy?” or “Will I get tangled in that rope?”. Moments like these are a wise distraction and must not be overlooked in underwater work, but a lack of overlap is typically the outcome. We didn’t know it at the time but we were not shooting overlapping images at at least two points.
Like all plans they never survive first contact with the enemy…or in this case, upon seeing the caravan a simple loop might not work. A barbecue had been added and the general state of collapse might lead to a chunk of old model refusing to marry up to the current version. For the couple of hundred images it took reshooting was a doddle, plus the roof was the perfect place to drop some scale bars onto. They can be used later as check measurements to verify the GPS scaling has worked. The Poppy was scanned too…small enough to nip round in a few hundred images.
Cut & Shut
The Wheelhouse was not scanned during the latest dives, nor was the helicopter. Neither are particularly large or complex but we are here to prove out working methods, not try and shoot everything.
With fundamentals like image data management in place finding the 2017 wheelhouse images was a doddle, so they were merged with the GPS referenced shots. All aligned first time, proving again we can mix and match images as long as nothing much moves, resulting in a separate chunk of 3D data in the workspace. Metashape Basic & Pro have several methods of chunk alignment, and as we have common images between two chunks we used Align by Camera. This method is very quick but does depend on accurate camera alignment…something now embedded in the process thanks to work with police forensics.
Thus we merged 2017 images with 2020 work and the wheelhouse dropped into place. The GPS referenced model acted as master and Metashape scaled the wheelhouse, applying estimated camera locations to the 2017 set. Accuracy needs proving out here (tape measure and divers needed) but the initial indications are positive. The scale bars placed on the caravan were within similar levels of variance between model and real world.
For those wanting a closer look at the ortho photo:
And the 3D model. If you look closely the blend between 2017 and 2020 model content can be seen around the base of the wheelhouse:
Frustratingly the existing images of the helicopter refused to merge. We think the primary reason for this are the sheer number of changes between first scan and today; the helicopter was scanned a few days after arriving and in the intervening years its condition will have changed, settling into the silt and the process of ageing means there is a lack of common points.
We can rescan it at another time after the UWIS system is returned to Pertti, knowing we have fixed its location in context.
Summary
Whilst we have gaps the GPS referencing of the cameras means we do necessarily need to have 100% overlap all the time. Each chunk knows where it is thanks to its reference. And with UWIS navigation we can guide any diver back to the missing point. Their swim pattern can be preset to gather the images needed.
The overall learning and experience this project has given us has been invaluable. This year more police work and a return to the SS Thistlegorm are planned, so the skills and knowledge will be applied.