It has become a bit of a tradition that I use this blog to make Mike Taylor of SV-POW! (and much other) fame a tiny bit jealous. By posting photos of the Museum für Naturkunde Berlin dinosaurs, for example a selfie with the skull of the Giraffatitan mount, or from other unusual perspectives – photos a normal visitor can’t ever take, and photos Mike (despite getting better access as a researcher) didn’t take during his short Berlin visits. His real-life job has given him far too little time to come visit. Still, the MfN Berlin dinosaurs have featured prominently on SV-POW! again and again. In fact a very special bone, the 8th cervical of Giraffatitan individual SII featured in the very first post there.
Mike, aside from being a very esteemed colleague in the same league as Eric Snively, Larry Witmer, Matt Wedel, Andy Farke, “Dino” Frey, Aki Watanabe, Michael Pittman, John Hutchinson, Viv Allen, ….. oh, jeez, I better stop before this becomes a ten page list of cool people in paleo!
Anyways, Mike is not “only” a really cool colleague and (American-style) friend, but also someone I personally trust in the way Germans trust their friends (which is on a totally different level than a US-style friend).
Given the affinity of Mike for the Giraffatitan‘s 8th cervical it is, I guess, especially suitable for making Mike all green-eyed. After all, while it was on display in 2005, today it has a new number (MB.R.2181.47 or MB_R_2181_47 in the computer-palatable version), rests in a wooden box in the bone cellar, inaccessible and hidden from view, and gathers dust – except for Wednesday two months ago. On that Wednesday, it was moved out by the MfN preparators to the hallway, and the sides of the wooden box were taken off, and the sand bags that stabilize the vert were taken away. For this:
Artist Alicja Kwade wanted high resolution scans of various bones, and for this very special occasion the museum OKed access. Alicja payed for the scanning of several vertebrae and ammonites by a professional surveying firm, matthiasgrote PLANUNGSBURO. Now, a lot of people have told me that they know some firm or other, and that said firm will quickly create perfect scans. A lot of people simply buy an expensive scanner, know roughly how to handle the software that comes with it, and them call themselves “experts”. Well, typically I have these conceited scanning “expert” for breakfast…… but not these guys! I was very impressed by their knowledge and experience. They know exactly what they are doing, know how to work to order (e.g., not creating a model at far too high a resolution, which means unnecessary cost), can do top-notch models if needed, brought a wide range of tools, all of which they knew exactly how to employ – it was great fun and quite informative to see them at work! And both the boss, Mr. Grote, and his two employees are very nice people, with whom I had some fun conversations.
Still, any such opportunity to scan difficult objects is a challenge for me, and this is especially true when someone else is scanning the same object at the same time! Can I scan as fast, as accurate, as detailed as them? Can I predict my data capture time and scan resolution and accuracy accurately? Does my data capture approach work at all? In short, can I hack it? I have recently become pretty cocky, given the success of the digiS bone cellar project‘s success, but that concerned rather simple bone shapes. This time, as I quickly saw, I was pitted against the elite of 3D scanning, and the specimens were of an entirely different level of complexity. Not that I expected the experts to beat the resolution of the model I was going to try and make with the scan they needed to do for Alicja – theirs would be for rapid prototyping on a CNC milling machine, and therefore of limited resolution, whereas mine would be aimed at way-more-then-enough for all science and exhibition uses I can currently imagine. But , knowing how scanning people tick, I was expecting them to additionally go for a top-notch scan anyways, going way beyond the ordered level of detail and resolution. And given the tools they brought and their expertise, I must admit that I was a bit afraid of working too quickly, taking too few photos and ending up with a model that has errors or big gaps, and compared badly to theirs.
In the end, as the photo above shows, along with an Artec Spider scanner they did bring the Big Gun – the Faro ScanArm with laser scanner. And they did go for a very high resolution and high accuracy scan. Which means that my best scan would have to measure up to a really excellent scans by them……. *gulp* I was quite a bit tempted to forego my usual happy-go-lucky high-speed scanning routine for a calmer, more thorough approach, maybe even using a tripod, simply to make sure that I drive quality up as high as I can. But then, the comparison is only fair if I stick to the same effort expended and use the same tools that I normally do!
So, they scanned with the Faro Scanarm and an Artec Spider scanner, and I used my trusty old Canon EOS 70D with a cheap LED ringlight. No tripod, no extended scan planning. Just my usual happy-go-lucky approach. Several vertebrae were set out on the work table in the Bone Cellar – not much room to work in, but sufficient for the artec scanner and my camera. The huge cervical 8 of Giraffatitan was moved to the hallway outside the bone cellar to allow better access with the Faro scanner, as can be seen in the photo above. And there it was that I went at it with my camera.
Overall, I took 754 images, the first 20 with scale bars placed all around the bone, the rest overview and close up photos without scale bars. Here’s one of the former:
The scale bars use the pre-made markers that come with Photoscan, so that the software can automatically detect them. This time it worked like a charm, saving me quite a bit of time. Matteo Belvedere is to be thanked for fighting with Photoshop to create the file from which we had the scales printed – thank you very much, Matteo! I used a bunch of 0.5 m scales, because scales half as long to slightly longer than the specimen you scan are best: they provide the least proportional error without causing extra work capturing them. And I must say that the resulting accuracy is pretty pleasing! Below you can see a screenshot showing the scale bars and their respective errors:
note that the average deviation between the scales, each 500 mm long, is less than 0.33 mm, i.e. less than 1/15th of a percent 😀 Photogrammery FTW!
After taking the scale bar photos I removed the scale bars from the bone. The same process – removing the photos with scale bars – I later repeated in the Photoscan project file, after alignment: I made the images unavailable for model creation. This way, they are there for scaling, but are ignored for construction of the dense point cloud, and do not litter the model. Because of this approach I can place scale bars ON the bone itself, instead of just around it, which gives me more flexibility. In some cases, like digitizing trackways, placing the scales on the specimen you wish to digitize is the only way to place them, so remembering the trick of using them for alignment and scaling but blocking them later is helpful.
The additional 734 photos fall into three distinct categories:
- images I took while “rastering” the bone
- images where I deliberately pointed the camera into recesses and at other difficult points
- overview images
The first category obviously is necessary to deliver a model that shows the entire bone at high resolution. I makes up about 1/3 of the total, because the photos need to overlap quite a bit. The second category makes up more than 1/3, not because I really needed that many (despite the plethora of deep, air-sac-caused depressions in the bone), but because I took way too many images, to make sure that I had enough to cover all the many nooks and crannies. Better having too many photos resulting in extraordinary calculation times than ending up with a model with unnecessary holes! Last but not least, the overview images are necessary to guarantee a good alignment of the other images. Yes, you can omit them, but if you take a series of images down one side of the bone and another up the other side, there is a high risk that your model will “warp” a bit. Overview photos keep this in check.
Rastering is best done by doing one set of photos with the camera pointed at the center of the specimen, then (for complex shapes like verts), another with the camera tilted left by ca. 30°, and another with the camera tilted right at 30°. Or up and down, depending on the shape of your specimen and how you place it. Or all of them – up, down, left, right….. and so on. Here, I made sure I used “straight”, “left” and “right”, as well as “up”. “down” images weren’t needed as a separate set, because of the geometry of the vertebra.
Then came the “recesses” part, which basically means pointing the camera at the midpoint if a hole, then moving it on the surface of an imaginary sphere but keeping it pointed at the same location. I did this for every single freakin’ depression….. *sigh*. I much prefer proper titanosaurs; they relocated their air sacs into the bone and have rather smooth outer bone surfaces. Much easier to digitize!
All in all, I spent 45 minutes and 21 seconds on photography, which does include a short 20 meter walk to a door and back to let some people in, as well as the time required to pick up and toss aside the scale bars. Divided by 754 images this means I took a photo every ~3.6 seconds. That may sound impressive, but it is actually slow work for me. Usually, I just aim the camera by eyeballing the brightness of the ring LED light on the bone. In this case, however, I felt the need for a more thorough approach, and used the 70D’s twistable touch live view screen to aim the camera and select the focus point. Usually, I achieve photo rates of 0.8/s, not .02777/s, but the live view screen makes shutter release slower, and the process of taping the screen each time to select the focus point and trigger the camera also is slower than simple blind point&shoot. Still, if I can’t easily go back and re-shoot a specimen, I’d rather spend more time and make sure I can guarantee a good model.
So, did all this effort give me a model I can be proud of? Can I hold my own against one of the best scanning crews out there? I can’t really judge, because I haven’t seen their models yet, but on the other hand I believe the results speak for themselves:
Yes, you read that correctly: the model has, in the highest resolution possible, some 484 million points in the dense cloud! Meshing a tiny part of itdelivers a 80+ million polygon mesh!
This is the full dense point cloud in all its glory! Note the hole at the bottom, where the vertebra rests on a plaster support made to fit. No way was anyone going to lift the vert up so I could take photos of its ventral side. It is way too heavy and fragile! We have very accommodating collections curators and managers at MfN, but lifting this bone is way outside anything they would ever consider – and rightly so! And a close-up – click for full size:
This area is less than 15 cm wide… oh yes, the resolution is amazing 🙂 Now let me show you the mesh…… below is a total of the dense cloud with a small part I meshed right away superimposed. Note that I did NOT yet clean the dense cloud at this stage, which is why there are ugly black rims on top and so on. The meshed area resulted in >80 million polygons, here decimated to 1 million.
and a zoom-in on the mesh (with slight smoothing):
yes, that hole you see is real! The bone really is that thin 🙂 I put two markers on the two sides of the neural arch that the mesh happened to cut. You can use Photoscan as a measuring tool by simply scaling a model creating markers and a scale bar from them, setting it to length 0, and checking the error – that’s the length of the scale bar (assuming you scaled your model correctly before). The thickness of the bone is really just
~4.569 mm! And despite the enormous size of the specimen, my happy-go-lucky model managed to keep the two sides consistently separate, except for the spot where there is a real-life hole in them, too:
So, overall, I am *very* pleased with my results! I haven’t seen the scans by Grote yet, so I can’ really say how I measure up against them, but I have once again been able to capture a very high detail model of a difficult object with simple, affordable and mobile equipment.
So, Mike, here it is now in all it’s glory – or should I say, in a small percentage of all its glory? As this is only a 74 million polygon model after clean-up, and if I ran this at ultra high resolution I’d expect it to have around 600 million. It is detailed enough, though, as it is….