YT YouTube2Text

YouTube2Text → Transcripts → Demystifying Gaussian Splatting for Archaeology and Museums | CAA Australasia Seminar | 18 June 2026

Demystifying Gaussian Splatting for Archaeology and Museums | CAA Australasia Seminar | 18 June 2026 — Transcript

by CAA Australasia · 16,977 words · 2,325 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Look, thanks for joining us this
0:01evening. So, um we are CAA. So, this is
0:05a seminar sitting under the computer
0:08applications and quantitative methods of
0:09archaeology society um the Australasian
0:12chapter um of which both myself and um
0:15Callum are committee members of. Um this
0:19is one of a series of seminars and
0:22panels and things like that we do. So,
0:24if you haven't come to them before,
0:25please do join. It's free. nothing
0:27stopping you joining CAA Australasia,
0:30even if you're not in Australasia. Um,
0:32so tonight we're going to really look at
0:36um some
0:38new tech, a new tech that I think is
0:40scaring some people, new tech that some
0:42people are excited by, um, some are not.
0:45Um, there's arguments over what it is,
0:47how it works. Um, I guess, but before we
0:50start off, um, I'm on Darking Jungland
0:53up here on the central coast of New
0:54South Wales. I I pay my respects to um
0:57the elders past and present um and I
1:02obviously um to everyone else from their
1:04own lands um where you're listening
1:07from. Um
1:10this is really a talk. It's a seminar in
1:12a way. So I'm going to be trying to do a
1:14bit of teaching, but I hope there will
1:16be a bit of room for discussion at the
1:19end and open discussion with everyone um
1:21here tonight. Um, but that's uh really
1:24what we're about and it's going to be a
1:26good chunk of case studies. So, um, I'll
1:30let Callum introduce himself when we get
1:32to his case study. So, he'll be doing a
1:34case study um, in the first part. U, but
1:37my name is Michael. I've, um, I've worn
1:40a bunch of hats in my career and right
1:42now I do two things. So, I um, oh, but
1:46before I start, there's um, all the
1:48links and stuff like that to join CAA.
1:50They'll be there again at the end.
1:52they'll be in the chat as well. Um, so
1:55yeah, we're going to go through intro to
1:57Gausian splatting, um, case studies from
1:59both of us, a little bit on processes
2:01and methodologies. This isn't a tech
2:03night. This isn't a a how-to every
2:06little bit of software night, so don't
2:07be afraid of that. Um, but it's it's
2:09really about demystifying and we'll have
2:11an open discussion at the end. So, I run
2:13Pedestal 3D if you don't know that. It's
2:15a it's a company I spun out of the uni
2:17sector and um now work across museums,
2:20archaeology um universities in mostly
2:23Australia. Um and you know pedestal
2:27isn't doing gausian splatting yet but
2:29like every piece of software related to
2:313D in the world we are looking into it.
2:33That should be coming soon. But I also
2:35run this which is a consultancy and I've
2:37been running this for quite a few years.
2:39So this is a consultancy where um
2:41traditionally I've done photography and
2:43photoggramometry and virtual reality and
2:45things like that in those same sectors.
2:47And now as you can see I've sort of
2:49penciled in gau and splatting down the
2:51bottom because over the last 6 months
2:53I've been looking at how to do that
2:54commercially um as well. Um so ready set
2:58splat this is the kind of stuff we're
2:59going to be talking about this evening.
3:00And these are some of the some of the
3:02works that I'll I'll show you some as
3:03examples and case studies. But all of
3:06these things in front of you are things
3:08well maybe not the one bottom right but
3:10have things that have scared me in the
3:11past the things that are you know as a
3:13photoggramometrist have said I can't
3:15scan that or how would I scan that um
3:18too hard or too expensive for some of
3:20them um and now we can and this is the
3:23bit that I hope you come out of tonight
3:25realizing that it's kind of liberating
3:27in a way it's liberated you know the way
3:28of saying and and my what sort of vision
3:31of what what I can do and what we can do
3:33in this sector as a new form of data. So
3:35it's pretty exciting. Um, so what is
3:38Let's start here. It's fun. So 3D gauian
3:41splatting is an explicit view dependent
3:43radiance field representation in which a
3:45scene is encoded as a set of anastropic
3:483D gausians each with position,
3:50co-variance, opacity, and spherical
3:53harmonic color optimized against posed
3:55input images via differentiable
3:57rasterization.
3:59So that's what it is.
4:02Yeah,
4:03hopefully you'll understand some of that
4:05later tonight, but in in a much better
4:07way to say it's, you know, we're we're
4:09representing a scene with millions of
4:10colored semi-transparent blobs, um, each
4:12with a position, shape, and view
4:14dependent color optimized to reproduce
4:16input photographs. That's a that's a bit
4:17of a friendlier way to say it. Or we
4:19could just say it's a cloud of colored
4:21fuzzy blobs that when you look at them
4:23together, it looks like a photo. And
4:24that's kind of what this is as a
4:27technology. That's how it works. That's
4:29what we're trying to do. Um, and you
4:31know, the the sort of the propeller head
4:34stuff at the top there, you will learn
4:36as you get into it. Just like you had to
4:38learn about polygons and and points and
4:40and textures and UV coordinates as you
4:42got into photoggramometry, here there's
4:44a new set of language, but in the end of
4:46the day, it's another representational
4:48methodology to use um to to show things.
4:51So, I'm going to start here. So, this is
4:53um my my um you know through line for
4:56tonight is that Bob Ross was the
4:58original Splatfather. You know, if you
5:00ever watched Bob Ross um videos, which I
5:02loved, by the way, which was him
5:04painting and and the way he would
5:06delicately use a house painting brush
5:09to, you know, just throw trees on the
5:11canvas and he would just throw bits of
5:12color up there and all of a sudden these
5:14trees would come forward. And it was so
5:16wonderful watching these paintings, you
5:18know, the way the way it would go from
5:20color and transform into these wonderful
5:22scenes that you see before you. So I
5:24sort of joke that that's what we're
5:25talking about here, right? That's what
5:27gauian splatting is. It's about
5:29representational rather than detail. Um,
5:32and I like to joke that instead of
5:33having happy little trees, which was one
5:35of his catch catch cries in all of his
5:38things, he painted in these wonderful
5:39happy little trees, we're now talking
5:41about happy little splats because you
5:43got to learn to love the splats. and
5:44we'll see what they are as we as we go
5:46through tonight. So I think I'll start
5:48off here and and this is the other sort
5:50of way I use a metaphor to talk to
5:53clients about this or to talk to
5:55colleagues about this because it's hard
5:57to understand and you know the first
5:59thing you'll always get is um which
6:02one's better you know Gausian splatting
6:04or photoggramometry so I put to you you
6:06know which one of these is more accurate
6:07and these are both contemporaries Gustav
6:10Cyabot um and Claude Monet were were
6:12both operating at the same time one was
6:14a realist one was an impressionist you
6:16know so the realist
6:17was really about capturing what was
6:19there, right? That's um an example. And
6:21depressionism though was about capturing
6:23the feeling. It was about capturing the
6:24mood. It was about capturing the light.
6:26It was a very different thing. But which
6:28is more accurate, you know. So that's a
6:31one question, but and when you get to
6:33nature, what about these? And again,
6:35it's a different Gustav um Corb
6:38in and Claude Monet on the right again
6:41with some natural scenes. And again like
6:43when we look at natural scenes from as a
6:46photoggramometrist they're things that
6:47scare us. When we look at buildings and
6:49rocks they're things that are familiar
6:51and friendly. But um again you know the
6:53realism is about capturing what's
6:55actually there. The impressionism is
6:57about capturing that experience of
6:58seeing if that makes sense. The you know
7:00what it actually really looks like
7:02rather than the way we can represent it
7:04with polygons.
7:06So what about these two? This is where
7:08we get into starting to look at some
7:10real world examples tonight. Um, one of
7:12these is a gausian splat and one is a
7:14photoggramometry model and they're both
7:16made from the same set of data. Um,
7:19here's another couple views, you know,
7:20when we look at that. So, here's a
7:21closeup of this um place that, you know,
7:24I didn't enter. Thankfully, it's
7:26asbestos was there. Um, but, you know,
7:28it sort of shows there's there's
7:30different rendering techniques at play.
7:33And you can see on the right it's made
7:34up of little blotches of color, whereas
7:36on the left it's made up of pixels. So
7:38it's solid versus, you know, ephemeral
7:41is is another way to think about it. So
7:43on this on this case, you know, you
7:45can't even read personnel there. You
7:46sort of can make it out on the left, but
7:48you know, again, it's a little bit
7:50blurry. So maybe neither side wins. But
7:53then I'll show you this, and this is the
7:55one where, you know, the rubber really
7:57hits the road when you start to look at
7:58the utility of this technology. Things
8:01that with photoggramometry are very,
8:03very difficult to show. For instance,
8:06the the tree on the on the um the
8:09chimney there, the the chainlink fence,
8:12as you can see, the vegetation itself
8:14are all all rendered very clearly like
8:16like a Monet, you know, they're they're
8:18splattered in and you can actually
8:20represent them really well. Whereas in
8:22photoggramometry, things are a bit, you
8:23know, harder um to resolve at that
8:26level. So, you know, that's that's sort
8:28of when we get down to it, yeah, you
8:30can't measure a mon. And this is the the
8:32other side of the main difference that
8:34realism which is photoggramometry is
8:36capturing the object. Um it's view
8:38dependent. So it's it's got a lot
8:40independent. Sorry. So once you've
8:41created the model you can look at it
8:42from any direction. Whereas gauian
8:44splatting needs views of all the
8:46directions to be able to create a model
8:48that works on every direction. So it's
8:49view dependent. Um you know photogometry
8:52all about the surface and measurement.
8:54It's very trustable nowadays or
8:56trustworthy. Um whereas impressionism is
8:58not. It's about the light. It's about
9:00the presence. It's beautiful and it's
9:01immersive, but it dissolves under
9:03analysis. You can't measure it. You you
9:06can measure bits of it, but it's not
9:08inherently trustworthy measurable. Um,
9:11so that's that's what will come up a few
9:13times tonight. But that lack of inherent
9:16measuring utility is actually it's not a
9:17flaw. That's what it is. That's
9:19splatting. Splatting is using a totally
9:21different type of technology that does
9:23stuff that um shows another modality
9:26that um we can't do otherwise. And and
9:29the joke is asking a splat to be a
9:30metric record is like asking a monet to
9:32be a building plane. It's just not the
9:34same thing. But the metaphor breaks, you
9:37know, the I can't, you know, thrash it
9:39too hard because, you know,
9:42impressionism was a rebellion against
9:43realism. Remember like there it was
9:45something that was trying to offer an
9:47alternative. I don't see splatting as
9:49that. I see splatting as a compliment.
9:51It's another tool. We will never replace
9:53photoggramometry um as a methodology.
9:56But now we have something that sits
9:58beside it. And in terms of, you know,
10:00archability, I think it's quite
10:01important to realize that splats are
10:04very new. So what does that mean from a
10:06museum and from a an archaeology point
10:10of view where we talk about data and the
10:12longevity of data and the truth of data?
10:15Um, it's a it's an interesting
10:17discussion to have with splats.
10:19So why does it matter? Why does it
10:20matter for all of you? Why does it
10:22matter for us here tonight? Because we
10:23are the CAA. Um, yeah, we capture sites,
10:26we capture objects, we capture moments.
10:28They're all things that we do already
10:30and we keep that data forever. So, as a
10:33new form of processing and as a new form
10:35of representation, it's an important
10:37thing to consider in our work. Um, I'll
10:40just quickly say what is a what is a
10:43Gaussian? Um, well, in a way it's a bell
10:45curve. It's a little blob. So, you know,
10:48Gaus himself um really defined a
10:51mathematical concept of a um a bell
10:54curve. mathematically. And so if you
10:56look on the right here, you can look at
10:58what looks like quite a normal scene,
10:59but the things in the foreground are
11:01just little splats of color. And if you
11:03look down the the, you know, sort of
11:05bottom middle, you'll start to see that
11:06they're all big ovals and they all have
11:09a color and they all have an edge. And
11:11then if you look really close at it
11:12bottom right, that's what we end up.
11:14That's what's making this scene. It's
11:15not polygons anymore. It's not pixels or
11:17voxels. It's splats. So that's the, you
11:20know, the splatting point. And we can
11:22call them Gaussians. We can call them
11:24splats. We can call the process gaussian
11:26splatting. Um I I'll come back to it
11:28later. I think there's a significant
11:30branding issue with all the language
11:32around this. But um that's what they
11:34are. And then a quick word on nerf. So
11:36you might go but isn't didn't we have
11:38nerf already? And Nerf's related but
11:41different. It's a neural radiance field
11:43and I won't be talking about that
11:44tonight. But the main thing about splats
11:46is that it worked out a way to encode
11:49that light view dependent information
11:52into points in space. So we could keep
11:54it and render it quickly. Whereas
11:55neurons neural radiance fields actually
11:58kept that in the model. So you could it
12:00was slow but you didn't end up with
12:02something that was ready to render on a
12:04graphics card. You had to then use the
12:06model to create an image from that um
12:08you know sort of neural network that was
12:10created. So, for now, Splatting won, but
12:12I'm not saying like it happened. Nerf,
12:14then Splat, and you know, that's it.
12:16Nerf's still going pretty strong, and I
12:18I don't think the race is over in
12:20anyway, but I just thought I'd mention
12:22that. Um, so tonight we're going to go
12:25through case studies on sites and
12:27objects. Um, and you know,
12:31photoggramometry wise, um, they're very
12:33different things. So, they're not just
12:35different subjects, they're different
12:36methodologies. when we talk about how we
12:38digitize and the museum and archaeology
12:40sectors again deal with both of these
12:42that we deal with the artifact itself
12:44and we deal with the context out there
12:45in the field. So you know the I think we
12:48know a lot about how we shoot those two
12:50things. Um you know photoggramometry
12:53does really well on sites but it can't
12:55manage reflections or transparency and
12:57very rarely do we capture what's around
12:59the place. We might capture a really
13:01nice view of the building but we don't
13:03get the city the building's in or we
13:04don't get the sky above the building.
13:06whereas Gaussian splitting does that
13:07really well. Captures the context around
13:09what's happening. Um and objects objects
13:12again photograph use really good
13:14objects. We've gotten really good at
13:15being able to do objects at a very
13:17accurate detailed you know repeatable
13:19manner. Um but proper physical B it's my
13:24boy. I'll fix it later buddy. Sorry.
13:28Sorry. It's my boy. Um thank you. So PBR
13:34is hard. So we can do PBR nowadays. So
13:36physical based rendering to to get
13:38shininess and things like that. Spiding
13:40just does it inherently. It captures
13:42light. It captures that representation
13:44of light. Um but you know things are
13:46view dependent. So a lot of the tricks
13:48we use in photoggramometry don't
13:50necessarily work um when we get to
13:52gausian splitting. And I'll unpack all
13:53of that this evening. But first let's
13:56let me hand to call um and let's listen
13:59to Callum about his experience as a
14:02digital archaeologist.
14:05I'm just getting myself set up.
14:15All right. Um, good evening everyone.
14:17Oh, that's not good. Skipped ahead in
14:19the presentation already. Um,
14:25there we go. Uh, good evening everyone.
14:28Uh I just like to start by acknowledging
14:30that I'm on Yukar country uh in Kool and
14:32Gata on the Gold Coast at the moment. Um
14:35pay my respect to uh elders past and
14:37present. Um tonight I'll be talking to
14:39you about some work I did last year up
14:41in the Northern Territory uh in a site
14:46uh called Nak Hill just next to the
14:49small aal community of Gambalana in
14:51Westland. So tonight I'm presenting uh
14:55all this with the permission of
14:56traditional owners and custodians of
14:57Inak Hill. Um
15:01and this is a project that has been done
15:03very much in collaboration with the
15:04community there.
15:07Um so a lot a lot of what I'm going to
15:09talk about tonight is the context of
15:10this project. Um because Michael's done
15:13such a good job of covering all the
15:14technical details. Um and I know he's
15:16going to go a lot more into that. I'm
15:18going to talk about some of the
15:19challenges of basically a digital
15:22workflow in a complex archaeological
15:26environment. One particularly focused
15:28around rock art. Um so Inac Hill is this
15:32fantastic environment. It's this big
15:33sandstone
15:35um rise o over the item land uh westland
15:39swamps. Um the the view on the screen
15:42there is from the 360 camera. So, it
15:45kind of gives you a bit of an idea of
15:47how high this sits above the landscape.
15:49Um, and the entire hill is covered in
15:51rock art. Um, Ben and Andrea, who I
15:54believe are listening in tonight, um,
15:56they're currently up there carrying on
15:59uh the project. Um, Andrea, Decra, um,
16:03and I think the the count is now well
16:05over 200 uh, rock art sites just on Inac
16:09Hill as as we're defining sites with the
16:11project. But really, it's a place where
16:15um you you turn a corner and there's
16:18rock art. Uh you just you can't stop
16:20finding it there. Uh which is a
16:22challenge cuz then you have to record
16:23all of it. That's the idea of this
16:25project is to actually record the entire
16:27hill in 3D. Um and I'll get a bit more
16:31into that soon. So just to give you a
16:33better idea of what it's like to work on
16:34Inak Hill, it is a difficult environment
16:37topographically. Uh it's a bit of a
16:40climb up to the get to the top of the
16:42hill. Um something that I I really felt
16:45last year when I was kind of carrying
16:47all my border and then all the
16:48electronic equipment up um is the kind
16:50of a challenge of doing digital
16:51workflows in a uh complex environment
16:54like this. You have to actually carry
16:56all that expensive equipment uh with you
16:58into the sites. Um so this presents a
17:01lot of challenges. You get this very
17:03variable environment up there. Uh
17:05there's lots of shade, but there's also
17:06a lot of bright light. Um so you're
17:08dealing with this heavily contrasted
17:10environment which from a photography
17:13perspective um for whatever reason
17:15photoggramometry gaussian splitting it
17:17can it can make things a bit trickier
17:19which is something I particularly
17:20encountered. Um so yeah you get lots of
17:24different environments up there. you
17:25know, you go from this basically really
17:26rocky kind of almost alien looking
17:28landscape to this densely vegetated uh
17:31spots as well and uh you know, presence
17:34of local wildlife which uh as we have
17:37found it could uh endanger the the
17:39recording process sometimes. Um but it's
17:43it's a it's a very vibrant cultural
17:45landscape. Um the rock art really
17:47reflects the environment that it's in in
17:49many ways.
17:51Um, so yeah, it's a tricky spot to do
17:54digital recording. Um, a lot of the rock
17:57art is in quite precarious locations to
18:00put it lightly. Um, I'm I'm not someone
18:04who's super comfortable with heights and
18:05uh this definitely challenged me on a
18:07few occasions when you're kind of having
18:08to, you know, stand on the edge of these
18:11big drops all around the place and uh,
18:14you know, try and balance the camera and
18:15not have it tip over and everything like
18:17that. Um, so yeah, you're dealing with a
18:20very complicated environment that has a
18:22lot of just physical challenges
18:23associated with it, but then a lot of
18:25challenges in then how you position
18:28cameras and uh laser scanners and all
18:30this other equipment. Um, but that's
18:33what makes it such a fascinating place
18:35is there's complexity to it.
18:38Um, the rock art itself is quite varied.
18:41Uh, that's one of the fantastic things
18:42about New Hill is you have all these
18:45different styles of rock art. um created
18:47across hundreds of generations over
18:50thousands of years. So you sometimes
18:53have these uh on the right there you can
18:54see one site. It's this incredibly
18:57almost chaotic kind of combination and
19:00layering of different um motifs of rock
19:03art. Um and then you have other ones
19:06which are much more isolated, much more
19:07distinct um but equally can be quite
19:11difficult to access. Some of them are
19:13kind of you you you look at them and
19:15think how how the hell did someone
19:16actually manage to put that up there?
19:18You know, there's no clear accessible
19:21point, which then becomes an issue for
19:22you, the one that has to record it. You
19:24have to actually kind of get the camera
19:25nice and close to get the detail you
19:27need to um actually record it to make a
19:303D model out of it or Gaussian splat. Um
19:33and then the the panel on the left there
19:35that that presents issues as well. you
19:36have these rock out panels which um half
19:41of it's in direct sunlight, half of it's
19:42in incredible darkness because of some
19:45kind of rock collapse or something like
19:47that. And the idea of what we want for
19:50this project is a cohesive kind of
19:52almost continuous 3D representation of
19:55not not just sites but the entire hill.
19:58So there's a real challenge there in
20:00consistently recording
20:02um imagery of these places that all
20:05meshes well together. Um so it's quite
20:08different from a nice controlled uh
20:11environment or your typical kind of uh
20:13open n uh environment. So yeah a lot of
20:17a lot of physical challenges that came
20:19with uh doing these recordings.
20:21Um so just talk about the the project
20:23generally on Trader Helon's decra. So
20:26it's a the whole point of doing this
20:28full scale digital recording of any hill
20:32is so we can actually understand how
20:35rock art has been positioned uh
20:37spatially and try and really get into
20:40the the depths of analyzing that. So, a
20:44lot of this is is about working with
20:45binning. The binning is is the name for
20:48the local people um up there. And um
20:52everything we do up on the hill, we were
20:54taking uh the guides with us to record
20:57sites, locate new sites, and they're
21:00often involved in the recording as well.
21:02Uh which is which is it makes such a
21:04more engaged experience cuz all this all
21:07this art has stories associated with it.
21:09Um,
21:11so it's it's a really really great
21:13engaging experience. So we're hopefully
21:17I don't know if I'm allowed to say, but
21:19there's a there's a paper coming out
21:20soon that's going to kind of deal with
21:22um how how the actual workflow for this
21:26project um happens. So, but it's
21:30basically the the idea is we have
21:32different scales of data collection and
21:35that includes stuff like drones, laser
21:37scans, photoggramometry, but then
21:39there's also additional things on top of
21:40that like um Q field which we're using
21:43for kind of tracking all the data
21:45collection process cuz you're dealing
21:47with so much data. You need to actually
21:49have a way of logistically orienting
21:52orientating how everyone's working
21:54across the different sites and what's
21:56being recorded, what needs still needs
21:57to be recorded. um you know oral history
22:00associated with it uh all this kind of
22:03stuff. It's it's a lot
22:05cohesive. Um so yeah the the ultimate
22:08goal really is you you collect all the
22:10data, you process it, you combine it and
22:13you end up with this incredibly rich
22:16data set which you can then start doing
22:18analysis of. You can you can start
22:20asking questions of why is this
22:22particular style of rock art here? Why
22:24why is it not there? All these kind of
22:26things. It's not my job to do that. Uh,
22:28which I'm kind of okay with because it's
22:30going to be very complicated, but I'm
22:32I'm really excited to see where the
22:34results of that go. Um, so another big
22:38part of this has been engaging the
22:40community. So, a big a big part of
22:43Australian archeological work is it
22:45being a reciprocal relationship with the
22:47abinal communities we work with. So um
22:51something has been the focus of my work
22:54particularly during my PhD was how do we
22:57actually take all this incredibly rich
22:59data that we're gathering for research
23:01purposes and actually make it accessible
23:05to the communities we're working with
23:06because often this is the case is it's
23:09definitely going to be an issue for
23:10gazium splatting it's been an issue with
23:12photoggramometry for a long time um
23:14something that that Michael's had to
23:16deal with that's that's I guess that
23:17kind of whole business model was making
23:19photograph guarantee model successful,
23:20right? Um, you know, a lot of the time
23:23these models have just been sitting on a
23:26computer somewhere. Um, a lot of the
23:28time they're done, you know, they're
23:31very high quality models that you can't
23:34run on the kind of hardware most people
23:38have. Um, so a big part of my work has
23:41been going, well, you know, the these
23:43are actually fun. You can use these
23:45models for things. You can make them
23:46engaging. You can give them back to
23:48community. So, um, one of the photos
23:50there, you can see Shawn, um, playing
23:52around with the, uh, Quest 3 VR headset
23:54that I took up the hill one day. Um, in
23:57that I actually had some of the I
23:58animated some of his art. Uh, Sean's an
24:01artist and, um, yeah, it it was it was
24:04great to kind of Everyone always loves a
24:07VR headset. You know, often when we're
24:09we're giving it to people, it's the
24:10first time they tried one on and um, it
24:13can really be a fun experience. Quite
24:15quite emotional at times as well. um
24:17depending on the context of the the
24:19cultural material being presented. So um
24:22yeah, we're always a big part of this
24:24project is looking at new ways we can
24:26actually deliver this very rich data set
24:28of in any hill back to the community. Um
24:31and I think there's a there's a number
24:33of ways that where we're looking at
24:34doing that. Um but there when you have
24:38this much data, there's so many
24:40different ways you can approach it.
24:42Um, so one of the the things I was
24:45testing up there was I I've been using
24:47the 360 camera. So I've been using the
24:50Insta 360 Pro 2. Um I I've been using
24:55using it mostly in a kind of uh like
24:57almost like Google Street View kind of
24:58sense, you know, like you can go around,
25:00you can collect a nice sequence of 360
25:03um images and then turn that into a bit
25:05of a a VR experience where you can put
25:06it on, get a high high fidelity view of
25:10an archaeological site or a cultural
25:12environment and kind of walk through
25:14that. um it's not very um interactive in
25:17a in a kind of game type sense, but um
25:20it's really about just actually getting
25:22a look and a feel for that environment.
25:24Um and you can be a still image or it
25:26can be a video. So that that's always a
25:27good good fun. But around the time we're
25:30heading up there, I thought, well, you
25:32know, I've been seeing a few people post
25:33online about using 360 cameras to make
25:36Gaussian splats. I thought, well, you
25:39know, I've got the camera up there
25:40anyway. I might as well test it out and
25:42see what we can do. Um so the the basic
25:46approach is that I I eventually settled
25:48on is um doing multiple kind of camera
25:52positions still still camera shots not
25:54video um around a site and
26:00uh typically on Inular Hills I was
26:02having to do exposure bracketing just to
26:05deal with the fact that a lot of these
26:07sites had extreme contrast of light and
26:10dark. Um, and if I just if I wasn't
26:13doing that bracketing, the results were
26:15well, I'm not going to show them. They
26:17were very very bad. Um, so,
26:21um, that that really adds to the
26:23process. I mean, I think I was settled
26:24on seven exposure brackets, so you're
26:27generating a lot of extra images doing
26:29that. A lot of extra data you have to
26:31store. Um, but the results the the
26:35results indicated that was the way to
26:36go. Um, so then you have to actually
26:38stitch the 360 images. Then somewhat
26:42counterintuitively, you have to un the
26:44process I use, you have to unstitch them
26:45again. Uh use Meshroom for that. And
26:48then once you've done that, you can
26:50actually put them into your splatting
26:52software to then generate the splat. Um
26:54I've also put some question marks there
26:56uh cuz we'll get to the whole issue of
26:58what you actually do with the splats
26:59after you've made it. Um
27:03so yeah, one the the the sites I was
27:06primarily testing at was um just down
27:08where that barrel is, you can see on the
27:10left hand side of the screen. This is
27:12this it's called um main gallery. So
27:14this is where tourists mostly go on in
27:18hill. It kind of the first big stopping
27:19point um for tours up the hill. Um I
27:24gave it a good go. It didn't It was my
27:26first test. I tried doing a video
27:28walkthrough with the 360 camera strapped
27:30to my back and it just turned out
27:33terribly. This is where I worked out the
27:35the contrast between light and dark just
27:39I I'd seen all these videos of people
27:40online. They just pull out their really
27:42cheap 360 camera. They walk in circles a
27:44couple of times and they get this
27:45amazing Gaussian splitter. I thought
27:47that looks great. That looks so easy.
27:49Not in this kind of context. It just
27:51does not work. you it it works in a nice
27:55evenly lit open environment but it does
27:58not work where you start getting this
28:00dynamic lighting. Um so that that that
28:03just didn't work out nicely at all. But
28:05then I then I shifted there was another
28:07site which I was more or less standing
28:09at looking down in this across the the
28:13south side of the hill and um the the
28:17results a bit better. I
28:19for this one I ended up combining it
28:22with uh as part of this whole project
28:25we're collecting photoggramometry data
28:26sets at every single site. So for this
28:29one I I what I found with the 360 camera
28:31even though it's very good it produces
28:338K 360 images when you actually compare
28:36that to the quality of DSLR image is
28:39actually a lot less kind of pixels um is
28:42a lot less resolution per area I guess
28:45per field of view. Um, and so the the
28:49results are actually not that not that
28:51good using just the 360c camera because
28:54the rocker I mean that that's the focus
28:56of this data collection is the rock art.
28:58You need it to really pop out of the
29:01image. You can't have it be this blurred
29:03mess. And I was finding with the um with
29:06the 360 camera, unless you even if you
29:08got it up really close, get lots of
29:11images that um really capture the
29:13rockart um it just it was really you
29:16have this kangaroo here. It's this very
29:18quite thin line. It's not the nice big
29:20blotchy bit of rockart. Um it was just
29:23blowing it out of existence. You
29:24couldn't even tell it was there. I found
29:27that combining the 360 data set with um
29:30the DSLR image set um more or less
29:33negated this issue entirely. Um I should
29:36be playing this video. Sorry. Um it's a
29:39very choppy one. This is just um a
29:41demonstration of what what the Gaussian
29:44splat result looked like uh using just
29:47the photoggramometry data set. Um so you
29:50actually get a fair bit of clarity on
29:51this quite fine uh rock hard. Anyway,
29:55this is a this is a relatively tall um
29:57panel. I think uh thinking back probably
30:01four four meters high maybe. So, you
30:03can't you can't get too high with the
30:05the 360c camera even though it had a
30:07pretty decent tripod. Um but and um same
30:11with the photoggramometry. You can only
30:12kind of hold the camera up so high to
30:15get close to the rock art. But um the
30:17the Gaussian spite actually did a quite
30:18a f fantastic job of keeping the detail
30:22um even on the more distant motifs on
30:24the panel. Um and the other the other
30:27thing I noticed and I'm sure Michael
30:28will get to this as well is um I found
30:30the Gaussian sweating process to be a
30:31little bit more forgiving uh than
30:33photoggramometry in terms of grabbing
30:35the little bits of extra little bits of
30:37extra um environmental context around uh
30:41that often gets cold I think in the
30:43photoggramometry process. Um,
30:46moving on. Uh, I did a bit of
30:49comparison. I I thought, you know, how
30:51does this actually compare to just
30:52photoggramometry? Um, and I I know
30:55Michael's going to get to this later,
30:56but
30:58it it it definitely struggles a bit
31:00compared to photoggramometry with hard
31:02surfaces. Um, it it just it seems to
31:06lose a bit of the crispness that comes
31:07with photoggramometry. Um, there's a bit
31:10of transparency to it. It becomes a bit
31:12ethereal. Um but I think I think that
31:15can be you know I think it it still does
31:18well given the right um data set it can
31:22get the job done. So yeah as a summary
31:26the FOV of 360s is great. It lets you do
31:29quite rapid image capture to some extent
31:31uh compared to a DSLR.
31:34Um but it it is coming at the expense of
31:37quality to some extent. Um, they do have
31:40a slightly, I think, extended workflow
31:42compared to a standard DSLR data set.
31:46Um,
31:48it doesn't matter. It's mostly
31:49batchable. It doesn't add too much extra
31:51time really. Um, and the the 360s do
31:55have some nice extra use cases. I like
31:57using them with sky boxes and this
31:59Google Street View type experience. It's
32:01a fairly minor addition, but it is fun
32:03and you can use them for both.
32:06Um,
32:09so yeah, what what what do you actually
32:10do with the splats? I we're going to
32:12talk about this more later, but a lot of
32:14a lot of the use cases at the moment are
32:16basically just kind of going back to the
32:17early days of um, you know, it's just
32:21walking simulators. That's that's a lot
32:23of what you see online at the moment.
32:25Um, you can walk around this nice static
32:27environment. It looks great, but um, you
32:30can't actually do too much in it. Um,
32:32and I think this big issue that it kind
32:35of fits into a couple of these is
32:38for so long existing workflows have been
32:40about mesh and texture. Um, you know, in
32:44in game development, it's all about your
32:46objects have a mesh, they have a
32:47texture. Gaussian sprites are completely
32:49different to that. And so all the tools
32:52built around mesh and texture, all the
32:54kind of workflows built around mesh and
32:55texture go out the window. Some of that
32:58is starting to change. So, you're
32:59getting uh tools that allow you to do
33:01kind of dynamic light changes on
33:03Gaussian splats, but it's um it's still
33:06got a very long way to go before you
33:08kind of we're going to reach parody if
33:10if we ever do. I think as Michael's
33:13alluded to, it's they are two different
33:16result. Then they shouldn't be treated
33:18as a race against each other. They're
33:21achieving different things at the end of
33:22the day. And one thing I I do want to
33:24touch on a little bit is um is actually
33:27is getting weird with Gausian splats
33:29because they are they are weird. They
33:31have this kind of ethereal quality to
33:33them and they can they can be both hyper
33:36realistic but a little a little odd.
33:38>> Um so I just wanted to play quickly um
33:42this little video from one one of my
33:44favorite um kind of Gaussian spite
33:47developers um Voxal Kai. Uh he's a
33:50Japanese developer and he he does really
33:53weird things with Gaussian splats and
33:54that's what I love about him. Um because
33:57they they are odd. They they can be
34:00hyper realistic. I I'd say they can
34:02produce environments to me in virtual
34:05reality that feel more realistic than
34:08anything else I've ever seen. But I
34:11don't think that's necessarily what the
34:12goal should be. I think realism can be a
34:15a false target in some way. um you know
34:18the technology we have available
34:20particularly in spaces like virtual
34:21reality it lets you do really weird and
34:24creative things that you can't do in
34:26real in reality like you know it yeah
34:29you can walk through the painting and
34:31check out the the internals of the
34:33painting you can't do that in reality
34:34why not why not make that the goal of
34:37the experience
34:38um but I think that it it's it's we
34:42should be cautious about just chasing
34:46um realism and that's I I I love the uh
34:51use of impressionism, Michael, that that
34:53you did because it it does speaks to
34:55that, right? You know, it that there is
34:57something special about something that
35:00it it represents more the feeling, the
35:02aura of something that it does, the
35:04reality of it.
35:05>> Um the happy the happy little splats.
35:08Exactly.
35:10>> Yeah. Yeah. And um look, it's it's
35:14a it's a struggle in archaeology, I
35:16think, because from a research
35:19perspective, we do we do need realism.
35:22From a scientific objective perspective,
35:24that is something we aspire to is to
35:26create the most accurate representation
35:29of something as possible. However, we
35:32have to remember at the end of the day
35:34that the things we're recording don't
35:37just exist within an objective universe.
35:39they exist within multiple different um
35:42ontologies, different cultural contexts
35:44which don't necessarily get reflected by
35:48an objective reality. And so I think we
35:50have to be open to being weird so that
35:52we can more accurately reflect those
35:55different cultural contexts.
35:57Um yeah, and I I'll wrap up there. I
36:00I'll just finish by saying this is being
36:02funded by Andrea's Decra project. um is
36:05being done in collaboration with Elac
36:06Arts, uh Griffith University. Um I I'll
36:11talk bit again a bit later when we get
36:13to discussion, but if you have any
36:14questions for me or anything like that,
36:16feel free to add me on LinkedIn or
36:18something like that and we can chat
36:19about it. But um I'll hand back to
36:21Michael.
36:22>> Oh, thanks Callum. That's really
36:24interesting. And look, I mean, is it
36:26more forgiving or is it not? And this
36:28is, you know, I do find it more
36:29forgiving for what you're talking about
36:31the environment around, but I find it
36:33brutally unforgiving in terms of
36:35lighting and object work. And I think
36:37that there's a real tension there. So,
36:39um, look, let's unpack that. Like, I've
36:41got I got about 4 hours worth of demos
36:44to show in 40 minutes. So, let's hammer
36:46it. Um, I'm 6 months in. Like, I I first
36:49did a splat 6 months ago. I shot this.
36:51This was part of Art Express last year
36:53where I decided to shoot um a pot with
36:56emu feathers. And as a
36:57photoggramometrist, you're probably
36:58shivering right now going, "Oh god,
37:00don't shoot that." Um but well, I have a
37:03method. I build a little perspect box to
37:05stop the emu feathers moving firstly. Um
37:07and you know, this was late last year
37:10and I ran it through in December. Um
37:12created this and jaw dropped. Like it's
37:15my my world is now changed, right? So I
37:17did emu feathers. Look at these
37:19beautiful emu feathers like every
37:21strand, every substrand, you know, right
37:23down to the riding inside the pot. This
37:25was new game for me, right? And this was
37:29a shot normally normal walkound method.
37:31You can see my point my my camera cloud
37:33there. I had a camera walking around
37:34taking photos with flash photography.
37:36That's the kind of result we can get.
37:38And that, you know, big deal. Um, this
37:41was one I did a little while ago and
37:42similar kind of thing but with a
37:44turntable. And I've struggled getting
37:46turntable stuff to work, but I think I
37:48cracked it a couple months ago now. Um,
37:50I'll show you this one. This is a
37:52beautiful sculpture down at in the Yaka
37:54um cultural collection down at
37:56Deniloquin um in New South Wales. And
37:58look at these feathers. Like look, this
38:01kind of thing that you know you could
38:03not do with photoggramometry just
38:05doesn't work. You know, they're
38:06liberating. This is liberating. And you
38:08can share things now that we couldn't
38:10share before. But if you look right up
38:12close, you have happy little splats.
38:14It's all just blurry things right up
38:16close and you'll learn to love it. You
38:17learn to love that feeling. Um, this one
38:20turned out surprisingly well. U, but
38:22what you'll find on some of these
38:23examples is if you look at the bottom,
38:25it's not there. So, this there's still
38:28some tricks with doing objects where you
38:29try and get the top, the bottom, all of
38:31it, it's really hard to stitch those.
38:34Right? So, that's one lesson that I've
38:35learned over the last six months that um
38:37getting down to something that you can
38:39see my lights reflecting off it, but
38:41actually getting it to work, you know,
38:42you're seeing the sort of internal
38:43splats going on there. That's kind of
38:45fun to end up in this weird world inside
38:48the pot, which is all the, you know,
38:50representations of those reflections um
38:53inside the pot. So, that's quite an
38:54interesting one, I think. Um and again,
38:57I'm shooting this normal
38:58photoggramometry method. Um let's show
39:01this one. This one's an interesting one.
39:02So I processed this only two days ago
39:04for another talk I did for for N
39:06yesterday. Um I shot these seven uh
39:10almost eight years ago. Um which is the
39:12Canandra fish fossils. I we spent 5 days
39:15in a shed shooting fossils.
39:18Basically we did 80 tons of fossils in 5
39:21days. I shot probably 40 pallets like
39:23this. So I pulled it out and I thought
39:25I'd I'd splat it. Why not? Um legacy
39:28data is a big one here. Legacy data is
39:30valuable again, right? So, if you didn't
39:32hold on to your raws and you didn't
39:34listen to all of us saying never get rid
39:35of your photos, well, too bad. But for
39:37me, I've got them. So, here we now have
39:40those rocks reprocessed and beautiful
39:42fish scales in the fossils, you know,
39:44amazing stuff. But this one, I'm showing
39:47it to you because of some errors. So,
39:48let's find an error. There's there's one
39:51right here where it actually goes
39:54transparent and you can sort of see it
39:56disappear. See how there I can see
39:58through this bit of wood all of a
39:59sudden. And if you zoom out, you can
40:01actually see the square. Like if I draw
40:03on the screen, you can see that there's
40:05a square right here, right? So that's a
40:07photo. And that's a photo that's exposed
40:09differently. And that bit of wood isn't
40:10in that photo. But what gausian
40:12splitting does is it tries to stay true
40:14to all photos. So if you have one photo
40:17that is disagrees with all the other
40:18photos, it does some really weird stuff
40:21to to stay true to that. So where, you
40:24know, it's made these wonderful little
40:25voronoid patterns that I find really
40:27odd. And there's another one over here.
40:29If you look at this um if there's the
40:31right angle I get. Yeah. Where it
40:32disappears just down here. So when I you
40:34know when I went back to the actual
40:36project which is here in reality capture
40:38that I did seven years ago. This and
40:41this are those two shots. So what I did
40:44on every pallet was I took a shot of the
40:45reference numbers as a reference for the
40:47photos and they ended up being included
40:49in my in my um animation sorry my
40:52alignment set but I hadn't put the bit
40:54of wood down yet. So they shouldn't be
40:56included. So, you know, to fix it in in
40:59gausian splatting now, I should be
41:00removing those photos. I should be
41:02finding the ones that don't agree with
41:03the rest and getting rid of them because
41:05it's unforgiving. Yeah, it will. Well,
41:07it's it's not unforgiving. It's too
41:09true. It's trying to create this, you
41:11know, representation that's multi- view
41:14relational that's true to all photos.
41:15So, if you include the photo, it's got
41:17to stay true to all photos. So, you
41:19know, those two are quite interesting
41:20ones compared to to the rest of the set.
41:22And that was a big lesson of mine um on
41:24a few of these bits of data. Um, I'm
41:26going to show you about boats. This is
41:28my holy grail or it's I have so many
41:30holy grails, but this is one of the holy
41:32grails I've had for years. I get asked
41:33to scan things like that. You know, it's
41:35a it's a boat model with little wires
41:37everywhere. And what do you do as a
41:38photographist? You you just panic,
41:40right? You go, how do I shoot that? I
41:42shoot it, but the ropes aren't going to
41:44work or the sails aren't going to work.
41:45You know, I got to I'd have to pay
41:47someone to model all those. It'll end up
41:49being very expensive. So, this year I've
41:51been doing some experiments. This was my
41:52first experiment. It's a model of the
41:54Yara that I shot in at the Maritime
41:57Museum and it showed up pretty good.
41:58Like did the ropes really good. Um some
42:01of the shiny bits not so good. I get all
42:03these holes in the side. I think that's
42:04what Cullen was talking about. The
42:06alignment is not so good. Especially
42:08where there's misalignment. You look at
42:10on certain angles, it looks kind of
42:11blurry. But generally, take a step back
42:14like an impressionist painting. Amazing.
42:15Look at that boat. Like couldn't it do
42:17that with photoggramometry? And that was
42:18a quick shoot. Um but what I learned was
42:21where I shadowed it. So, where I was in
42:23front of the light, there's another one
42:25there. You can see this dark square
42:26going on, right? Because I was trying to
42:28stay true to all those photos. So, that
42:30that was a good lesson. But I, you know,
42:31I tried again with um Parah. So, this is
42:34HMAS Paramea, similar kind of size boat
42:36um out at the um cultural center in
42:40Paramea Council. Um I shot this one with
42:43a ring flash. So, I thought, well, if I
42:44don't get in front of the light, I can't
42:46throw a shadow. So therefore, but it it
42:49it's it's a success, but gives you this
42:51really weird sort of, you know,
42:53spotlighty look, right? Because it's
42:54saying true to all the photos and ring
42:56flash falls off very quickly when you're
42:58doing shooting. Um, so I called it a
43:01success, but you know, I thought hard
43:02about this and that's where I ended up.
43:04Um, and I I'll show you the methodology
43:06shot first. I'm like, well, if shadows
43:07are a problem and lighting is a problem,
43:09why don't I've done this before. Don't
43:11light the model. This was my result. See
43:13my flash? I I lit the room. So, you
43:16light the room. Basically, you put
43:18everything in a big soft box, so
43:20everything's light. And then you're very
43:21careful to to shoot that model. And we
43:24have success. We have gorgeous success.
43:26So, this was um in at the Maritime
43:28Museum um photo studio. You can see the
43:31wonderful place that I shot this. And
43:33you can see how I've perched it up onto
43:35some perspects um blocks there so I
43:37could get shots from underneath. Like I
43:39said before, you can't, you know, mask
43:40that necessarily and do it that easily.
43:43Um, and and now I've got a result, you
43:46know, just wow. This took 15 minutes to
43:49shoot. This took half an hour to process
43:52this from a from an approach. Look,
43:54there's I it's not entirely true.
43:56There's a there's a devil in the detail
43:58of getting a good alignment. So, getting
44:00a perfect alignment that looks right
44:01from every angle. But this is a game
44:03changer. Like, look at that. Like, you
44:05know, from a photogometry approach, we
44:07wouldn't even try and do something like
44:09this. Um, now quick, easy. Um, but
44:12especially things that thin things like
44:14the sails as well just show up
44:16beautifully. Um, and some of these it's
44:18not just showing up beautifully. It's
44:20like which one is it? I did a few boats
44:22that afternoon. Not that one. That's
44:24another big boat.
44:26What about that one? Yeah. Yeah, this
44:29one. Um,
44:32so it's cuz it's view dependent. It
44:36actually shows translucency. And if you
44:38look at this really carefully, you can
44:39see the sail behind this sail. You can
44:42see the translucency that's actually
44:44going on there because it stays true to
44:46all the photos. So this is a this is the
44:48whole point of light fields that um is a
44:51game changer. The way I see it is that
44:53you know you're you're creating this
44:55point cloud and these gorians are being
44:57used to be right from all the photos you
44:59took. So rather than take your photos,
45:01make a mesh, put color on it, and you
45:04you've made something very hard, very,
45:06you know, representational of that
45:07thing. This is not. This is liberating.
45:09Um, shooting's hard though. So, you
45:11know, like I I sort of got it down to
45:13this point where, you know, shooting
45:15soft light everywhere. Beautiful
45:16results. And, you know, I I I can't wait
45:19to do more. I I'm loving some of the
45:21stuff I'm getting out of that. Um, I
45:23also tried it on some other projects.
45:24Let me show you this one. This hasn't
45:26been released yet, but I shot this last
45:28year. It nearly killed me. It was a six-
45:30day job in at state library. Um so I got
45:33asked to shoot the McCory collector's
45:35chest that's a very um old and very
45:37important piece of um early colonial
45:40history and it's got three lines to um
45:44indigenous and all kinds of things going
45:45on. So, I shot this. I'll show it in
45:47pedestal just to show off that. Um, we
45:49ended up I got asked to shoot this,
45:52right? Every bit, every drawer, every
45:54shelf, every bit of, you know, frame,
45:57all of it, right? So, I shot that up,
45:59down, everything else using perspect
46:01glass. It was not an easy shoot. If you
46:03look at this from a photogometry model,
46:05good result. Everyone's happy with the
46:07result. It's I'm happy with the result.
46:09You know, I I did what I could, but I
46:11thought it was an ideal opportunity to
46:13take that data and process it. So, if
46:16you actually look at this one as a
46:17photoggramometry model
46:21and loading's my only enemy tonight,
46:24it's doing okay. It came up good. This
46:26was the hardest of all of them because
46:27look, I had bugs with needles and thin,
46:31you know, cicada wings, like all this
46:33stuff. Photoggramometry doesn't do that
46:35stuff well, right? And I I remember
46:37telling at the time like, "Oh, get a
46:38good representation of it. like from a
46:40distance it will look like it but you go
46:42up close it's going to look blobby and
46:43blurry and that's what it does right but
46:45again it's you know doing the job for
46:47what it was for so I I took that and you
46:50can see the amount of work I did down
46:51bottom bottom left there on the amount
46:53of photos and you know very careful
46:55photography um created a splat and the
46:58splat you know so this was a chro splat
47:01I didn't even know about splats then I
47:02hadn't even started on splats but I
47:04reprocessed it and what like it's some
47:07of it is actually lovely Like if you
47:09look in this bit here, the shininess of
47:12these beetles that photoggramometry
47:14fails hard at. Even a PBR workflow would
47:17fail hard because it's it's this um you
47:20know fluorescent kind of you know
47:22subluminance. I don't even know how to
47:24describe it. But we know that kind of
47:26shiny pearlesence I think is the right
47:28word. Um that you can now capture with
47:31this tech. But if you look at the hard
47:32surfaces like Callum said, the edges it
47:34doesn't, you know, there's holes in it.
47:36It's not quite as good. So, we're right
47:38in this middle bit, but you know, my
47:40favorite is the Cicada wings. They're
47:41just gorgeous. They look absolutely
47:43amazing like up close, you know, as a
47:45splat um on this kind of thing. So, that
47:48I mean, that's that was a good test of
47:49old data. And if I shot something again
47:51for that, I'd shoot it in slightly
47:52different way. But what it did work on
47:54was the the birds. The birds are just
47:56stunning. Like stunning. but also that
47:59look at the the feathers and the
48:01shimmeiness of the feathers that you get
48:03on things like this and the way it
48:04captures all of the details and the and
48:06the tiny little details going on.
48:08Amazing tech, right? Not measurable. No,
48:11not animatable in the same way.
48:13Different kind of thing underneath. Look
48:15at it. It's a mess, right? So, you can't
48:17really do all of that, but pretty cool
48:18tech. Um, here's a hard fail. So, I also
48:21shot this hoof. This hoof I just shot
48:24normally, right? I shot it as I would
48:25normally shoot photoggramometry at the
48:27time. I put it on a turntable. I shot it
48:29with a bunch of different rings. Um, I
48:31ended up with a hoof. Great. It looks
48:33like a hoof, right? It's a pretty good
48:35photogometry model. Um, I threw that
48:38into a splat workflow
48:40and I call this a failure. And this is
48:42where I first started to learn about the
48:44tolerance of um, splatting. So, firstly,
48:47it's very shimmery, but it's more at
48:49certain angles. When you look at this at
48:51certain angles, let's see if I can find
48:52one. Yeah, there. See this bit down
48:55here? That's where I masked it because I
48:57had to mask off the support. It fails
49:00altogether. It can't handle that
49:01discrepancy, view dependent discrepancy
49:03between this is masked here and not
49:05masked there. So, it fails. And you
49:07know, you can get some really quite poor
49:09results, you know, when you look at some
49:11areas close. Um, yeah, there like things
49:14flashing in and out like this one. Yeah,
49:16not great. That one I then went back and
49:19I I at the same time as all this, I've
49:21been learning PBR a bit more and Okay.
49:25All right, load. And I went back and ran
49:28it through my PBR workflow, and it's a
49:30much better utility for something like
49:32this. That's a much nicer scan, which is
49:34back to normal mesh, but a normal mesh
49:36with a roughness channel that, you know,
49:38really, you know, did that same kind of
49:40work for the shimmery, but a different
49:42kind of result. So, does splatting
49:44always do surface better? No. Like, it
49:47does some surfaces better, like
49:48pearlescence. Some surfaces we can get a
49:50good PBR result um off some things.
49:53Here's one I did recently. This is um a
49:56job I do every year on um fashion. So
49:59it's um here's some puppets. You know,
50:01again, things that are hard like
50:03photoggramometry wise to do to shoot
50:05puppets. Um you know, these these are
50:08are pretty cool. Like there's a lot of
50:09floaters on this one. There's a lot of
50:10discrepancy, but again, like gives you a
50:12nice feel. Um and it's a little weird,
50:15yes, but I've got these kind of hanging
50:17puppets that that really get that
50:18feeling across. And the good old rhino
50:20is my um my favorite which is one that
50:24was in the gallery floor and
50:25photoggramometry would fail at this and
50:26fail hard because it's got all this wool
50:29and fluffiness and fuzz you know being
50:31able to do fuzz that's liberating you
50:33know to get fuzz you know in what we can
50:36scan now and 3D scan fuzz and fluff and
50:39things like that that have always been a
50:41no-go. They've been things that you get
50:42cropped in photoggramometry or just you
50:44know don't show up at all. So, you know,
50:47so cool. And you can see the way I shot
50:48it down the bottom. That's my same kind
50:50of approach of light the room, not the
50:52object. And that's a, you know, a key
50:54lesson that if you take away tonight,
50:56take that lesson away because, you know,
50:58nice soft lighting, nice non-variable
51:00lighting. But that, I mean, I've been
51:02doing that project for years. It's my
51:03next Redux because I've got years worth
51:05of fashion that I is all sitting on the
51:08shelf as as legacy data that I'm going
51:10to redux into um splats and see see
51:13where it works and learn more. Um, and
51:16that's what I then did to Art Express.
51:17So, I do Art Express every year. It's a
51:20um a project where I do high school
51:22artworks. Um, I have seven years of
51:25data. I have about 180,000 raw photos
51:29I've taken. So, I started reduxing them.
51:31Why not? I went through some of my scans
51:33and this is was never shot for this
51:36purpose is just ridiculously cool,
51:38right? the the individual fluff that is
51:41now evident on a on a sculpture like
51:43this that I made a photoggramometry mesh
51:45of it and it was good. It would read as
51:46the object, but the fluff would never
51:49read. So, that one's a beautiful one.
51:51Some of the other ones that showed out
51:52really well was um uh ones like this
51:56that's got this um sort of, you know,
51:58translucency and metalness kind of
52:00characteristic to it that came up really
52:02really well. Again, not shot for this
52:04purpose. Shot for a different purpose.
52:06There were some like this one that
52:08actually failed when I shot it. I shot
52:09this for photoggramometry. Um, and it
52:12failed. It failed hard. It didn't work.
52:13There were holes all over it. I actually
52:14had to abandon it. Um, went back and
52:17created this model that is just amazing,
52:20you know, for for what it is. It's a
52:22it's very um it's a it's a straight
52:25jacket talking about um, you know, the
52:27the students mental state at the time.
52:30It's very teenage angst some of these
52:32artworks, which is just awesome. But um
52:34this one was a yeah an Art Express
52:36winner that year that now has a new life
52:38and has a new way to look at it. Um but
52:40and then some of the other ones that
52:41really stunned me that um I I shot this
52:45at the time. I'm I'm brave. I'll shoot
52:47it. Why not? Like let's try it.
52:48Experiment, experiment, experiment. You
52:50don't know with photoggramometry if it's
52:51going to work or not until you try. Um
52:53and I shot this one again in a perspect
52:55box with fishing line hanging these, you
52:58know, knitted jellyfish. It's such a
53:01cool thing. But now look at it. It's got
53:03a new life. It's like the fishing line
53:04itself like showing how this was built
53:06and how it was suspended and it's got
53:08this aireriness and you know, yes,
53:10there's sort of blurriness to the to the
53:13um perspects there, but it's just it's a
53:15wonderful thing. Um and you know, I've
53:18done so many now um that I I just want
53:21to keep reduxing work um from the past
53:24few years. But all of these get
53:25launched. You know, things like this um
53:27are going to get launched in the next
53:28few weeks when the new Art Express comes
53:30out in um the week and a half's time.
53:33So, you know, um that launches soon. Um
53:37so, that's objects. Whoa. Yeah. Okay.
53:40Keep going at this speed. That's what I
53:41have to do. So, let's move to sites. So,
53:44that's objects, turntables, walk around.
53:46You know, similar kind of thing. Some
53:48things work, some things don't. Um I I
53:50will show you a quick couple failures
53:52first from Art Express. This is a
53:53failure. This is a hard failure. This
53:55kind of splats everywhere. This mess of
53:58things in the air. I actually um worked
54:00out why it failed and it failed. Here's
54:03the actual photoggramometry model. It
54:05failed. And my belief of why it failed
54:08is if you actually look at um the the
54:11shots I have down there, they've got the
54:12white um lighting tent in the shot and
54:15that white lighting tent doesn't move
54:17and the object moves. So splatting is
54:19trying to stay true to all that. So it
54:20has to put these little blobs in the
54:22air. and you know, I'm going to redo
54:24this one by masking that out and I think
54:25I'll succeed. So, I've learned that kind
54:27of the void is important again when
54:29you're doing turntable work and making
54:31sure you don't have things that overlap
54:33what you're working on in rotational
54:35senses because it's unforgiving. It has
54:38to be true to every photo. So, that's
54:40both, you know, forgiving and
54:42unforgiving if you look at that in in an
54:44operational sense. And then the other
54:46one that's um uh quite an interesting
54:48failure was things with deep recesses
54:50like this is a little um Japanese um you
54:55know hut that was done by a student made
54:57in balsa wood and it kind of looks nice
54:59on the outside but inside it fails.
55:01Inside it fails because I didn't have
55:03enough points of view of this. I only
55:05had it from one angle and if you only
55:07have it from one angle splatting is even
55:09it's a lot worse than photoggramometry
55:11at being able to do a surface. you know,
55:13it's a lot less tolerant to knowing
55:14what's actually in there. So, you can
55:16see it's all this this mess of
55:17blurriness going on. So, another failure
55:19that wasn't included. Um, okay, let's
55:22move on. Let's move on to the engine
55:23room of the vampire. This worked. I shot
55:26this 3 four years ago. I It's It was a
55:30horribly hostile place to work. There's
55:32asbestos everywhere. Bonded asbestos.
55:34Um, it was loudspace. It was oxygen
55:38space. All kinds of things on the HMS
55:40Vampire. You can see my um point cloud
55:43um down there. There's these little um
55:45circles everywhere which are my laser
55:47scan points. So, I brought in an Artech
55:49laser scanner. Um I really scanned the
55:52heck out of this place. And that bottom
55:54right is my photoggramometry model,
55:55which is good, right? It it came up
55:57well, but it has the problem Callum was
56:00talking about. It's too big to publish.
56:01It's got too many textures and too much
56:03of this and too much of that. So, this
56:05one 3 years on, I reprocessed as a
56:08splat. What a just this was Yeah, this
56:11was probably the second or third thing I
56:12did in December which just my jaw
56:15dropped. Like all of a sudden I've got
56:17this engine room and if I look up close
56:19like all the details there like it's I'd
56:22shot it well enough and I'd lit it well
56:24enough that all of this becomes very
56:26readable. If you look too close, yes,
56:28it's all splats, but what what I love
56:30about it is the the depths. This was the
56:33engine rooms are pretty full on place.
56:35So there's, you know, in the depths of
56:36places over there that I didn't even go,
56:39you can go and look down into the next
56:40level down and start to see that
56:42peripheral. And I think that's what you
56:44were talking about, Callum, that um, you
56:46know, unforgivingness. So there was and
56:48no way photogometry would ever render
56:50that bit of this scene. I didn't take
56:51any photos of it. It was background
56:54information. Yet splats stay true and
56:56you get this wonderful thing. This one's
56:58got a lot of floaters in the air.
57:00Floaters are how splats deal with
57:02discrepancy. So, they're not necessarily
57:04a bad thing. They're not necessarily
57:06something to love, though. We go in and
57:07we clean them up, and we'll talk about
57:09that um in a minute. But that one's kind
57:12of cool. This one I showed you before.
57:13So, this one, just to um go into a
57:16little bit more depth, I I did a quick
57:18shoot. It had a wire fence around it. I
57:20flew it. It's drone only. So, I flew my
57:22Enterprise Mavic, my Mavic 3E. Um I flew
57:25it up and down the hallways carefully. I
57:28forgot my prop guards that day, so I was
57:30very careful. So I flew it um inside the
57:32fence between the fence and the building
57:34you know all all kinds of places and
57:36ended up getting quite an impressive
57:38scene. So that was you know the first
57:41time seriously I tried to do this out in
57:43the fields and I think it worked. I mean
57:45like look at the chain link fence. This
57:46is something that is very splat
57:48renderable because it's a lot of long
57:51blurry things, right? If you look if you
57:53if you from an impressionist point of
57:54view a chainlink fence is just a lot of
57:56criss-crossing blurriness. So a splat
57:58does that. If you look at it that's what
57:59it is. a lot of criss-crossing
58:01blurriness. Looks great. Things like the
58:03um the the um aerial the antenna came up
58:07really well. Things like this, you know,
58:09bit of a tree came up really well. But
58:11if you look close to anything, yeah,
58:12it's blurry splats. Um so, you got to
58:16love them. But yeah, even the pegs
58:18hanging up here I thought was really
58:20interesting. You know, sort of the way
58:22it could render all these things staying
58:24true to it. So, that was a single drone
58:26flight. Um, and that one I also then
58:28processed into a a mesh. And I'm doing
58:31this more and more. And that's the punch
58:33line of where we're going to get to the
58:34end tonight is shoot for both. Learn how
58:36to shoot for both. Why pick one? You
58:38know, the the right answer is shoot for
58:40both. So, this feels a lot more solid.
58:42When you look at it, it looks like a
58:44photogometry mesh, but not just looks
58:46like one. It is solid. It's measurable.
58:49Yeah, we all know this about
58:50photoggramometry now. We can calibrate
58:51scale and we can get very accurate
58:53measurements, accurate drawings, things
58:55like that. It's a different utility,
58:57right? So, that's the the same data set
59:00um as that one. I also tried Colum's
59:03approach with a 360 camera. I bought
59:05myself an Insta 360 X5, the new ones,
59:08and I shot this rather quickly. I took I
59:12took 10 minutes and I walked around um
59:15at this cemetery, which was just down
59:16the road. And again, it's not bad for
59:19what it is. It's not bad. If you go try
59:20and look up close to any of those
59:22gravestones, it's garbage. You can't
59:23read anything. It's a big blurry mess,
59:25but do you get the feel of this
59:27abandoned countryside in the middle of
59:28the bush? Absolutely. Yeah, it's it's
59:30wonderfully atmospheric and the way it
59:32it paints the trees and and gives you
59:34all that. And this one was literally,
59:37you look at my point cloud was I just
59:38walked around the perimeter and then I
59:40walked a zigzag and it was all on video
59:42on a pole. I just walked around, you
59:44know, I pulled the shots out of that
59:46video. You know, from that you sort of
59:48cut them into slices so you unslice them
59:50like Callum was referring to. And it I I
59:53call it a semi fail. I wouldn't go and
59:55publish that. I probably if I was doing
59:56commercially, I wouldn't want to charge
59:58for it because it wasn't that great. But
59:59I learned a lot about sort of dependency
1:00:01of um cameras.
1:00:03More and more now though, I'm doing it
1:00:05purposefully. So purposefully on some of
1:00:07the shoots. So I do a lot of work in
1:00:08dryd do. I do a lot of work with boats
1:00:10like the the um doken as you can see
1:00:12there in dry do. Um you can see the kind
1:00:14of point clouds that sorry camera
1:00:16coverage that I'll do when I'm on site.
1:00:18So I use drones, I use ground photos, I
1:00:21do some significant um coverage. So
1:00:24something like the this now I'm
1:00:26delivering as a mesh like that's the job
1:00:28but I get the context and I think this
1:00:31is the key for archaeology the context
1:00:33where it was on that day you know how it
1:00:36was positioned in the dry dock um the
1:00:38kind of work that was being done on it.
1:00:40I get a context that um then goes along
1:00:43with the mesh or goes along with the
1:00:45data. So you can have both. you can have
1:00:47that context. Um, you can shoot for
1:00:50both. So, this is the same, you know,
1:00:52data that I used to make the mesh as
1:00:54well. Um, but you get wonderful scenes
1:00:56like this, which is that, you know, boat
1:00:58in the the dry dock that it was. And
1:01:00yeah, up close like black things kind of
1:01:02look blurry and shimmery. Um, again,
1:01:06it's weird. I like that column. It's
1:01:08weird. There's a weirdness to it. But
1:01:10yeah, I love the weird. you know,
1:01:12there's something so real about it as
1:01:14well that is just, you know, something
1:01:16else um when it comes down to it. Um I
1:01:19reprocessed some of my other ones from
1:01:20the last couple years because I've done
1:01:22a lot of work in dry do combination of
1:01:24drone and ground and some of them are
1:01:26just wonderful. This is a very foggy day
1:01:28and you can really feel the fog of that
1:01:30day. Um you can sort of really get the
1:01:32position of where it was in dry dog.
1:01:34This is at no in North Sydney. Um I
1:01:37didn't clean this one up. I just left
1:01:38all the floaters and splats there
1:01:40because it was a test. But it does
1:01:41things like, you know, this hatch was a
1:01:43nightmare in photoggramometry. I had to
1:01:45do all this clean up and smoothing and
1:01:46rebuild all the windows and and rebuild
1:01:49all these hatches. You know, there's a
1:01:50nightmare. Um, you know, it does it
1:01:53effortlessly. So, it really gets that
1:01:55kind of, you know, wonderful um crisp.
1:01:57You know, the ropes are really crisp on
1:01:59this one that I actually built in
1:02:00photog.
1:02:02Now, you know, I'm really looking at
1:02:04using the model for the hard bits and
1:02:06the splats for the non-h hard bits, if
1:02:09that makes sense. Um, to is an
1:02:12interesting one is in the car park. I've
1:02:13been shooting this boat for years.
1:02:15Here's a really interesting snapshot of
1:02:17how it was that day. Um, again, it's
1:02:20just a quick I'm making model I've made
1:02:22models out of this. It's on display now
1:02:23and we 3D printed it. Um, really, really
1:02:26good. But in my in my models, I'll end
1:02:29up painting out all these supports. So,
1:02:30I'll cut them out from the
1:02:32photoggramometry mesh. I'll end up
1:02:34projecting, repainting the color to make
1:02:36sure they're not there. You know, you
1:02:38can't do that with splats. It doesn't
1:02:40have that pipeline yet, right? So, the
1:02:42pipeline for for meshes, you know, to be
1:02:45able to do that post cleanup, there's so
1:02:46many tools. There's so many options.
1:02:503D software is so mature that I can
1:02:53absolutely redo things like that. In
1:02:54splats, I can't. I got to live with it.
1:02:56That's what it is. So, it's great as a
1:02:58snapshot. Um, you know, my final mesh, I
1:03:00rebuilt all the boards that are missing
1:03:02and things like that. So, you know, this
1:03:04doesn't work as well for that. But the
1:03:05good old endeavor, look at this one.
1:03:07This was a bad day. I waited 3 hours
1:03:09this day. I I twiddled my thumbs for 3
1:03:12hours. It was going to rain. It was
1:03:13windy. I was waiting for the scaffolding
1:03:15to come down. And I did shoot Endeavor.
1:03:17I ended up getting half an hour to shoot
1:03:19this. And I shot a very quick um
1:03:21photogometry set at the base of
1:03:23Endeavor. So, this is the full size
1:03:24Endeavor in dry do. And then I I did a
1:03:27few passes with my drone to try and get
1:03:29the rigging. But look at it. It's just a
1:03:31mess. This is kind of if if you've ever
1:03:34tried to do this kind of work before,
1:03:35especially in the wind, this is not
1:03:37abnormal, right, for photoggramometry.
1:03:39Um so I just thought recently, why don't
1:03:41I'll give it a go. I did these boats.
1:03:43The boats worked. Maybe that'll work. So
1:03:46I grabbed it. I threw it in. Just
1:03:49stunning.
1:03:51Like the the ropes at the top, they're
1:03:52invisible because they were literally
1:03:54flapping in the wind, right? So they're
1:03:56moving and they they just didn't resolve
1:03:58in any way, sense or form. But down the
1:04:02bottom, beautiful. What a beautiful
1:04:03snapshot of the endeavor. So you know,
1:04:05it's given me
1:04:08bravery, I guess, to now approach things
1:04:10like this that I used to be afraid of
1:04:12this as a photoggramometrist that as a
1:04:14subject. It's a, you know, 100 m, well,
1:04:16not quite 100 m. It's a 70 m boat in the
1:04:19wind with ropes. It's a scary subject,
1:04:22right? It's a scary, you know, thing to
1:04:23try and scan. Now I'm kind of liberated
1:04:26to a degree. Um, so I kept going through
1:04:29old examples and then so thankfully as a
1:04:31client, they're a really good client.
1:04:33They they were brave enough to trust me
1:04:35to to have a go at this and we took the
1:04:37research boat up to the Hawkbury. Um, we
1:04:39we powered out to HMS Paramea. Um, and I
1:04:43shot this. So this is a splat. This is a
1:04:45wither drone. I we couldn't get closer
1:04:48than 150 m or we would have ground the
1:04:50boat. You couldn't get on the ground.
1:04:51There was too much mud everywhere. Um,
1:04:53so I very carefully flew my drone at
1:04:56distance from a boat. It was a scary
1:04:58day. I did four flights, I think. Um,
1:05:01and ended up capturing this, which is
1:05:03the HMAS Paramea wreck. So, it's our
1:05:05first Navy ship. Um, that's wrecked up
1:05:08on the Hawkbury right near Mooney
1:05:10Mooney, where you probably go over the
1:05:11Hawkbury if you live on the central
1:05:13coast like me. Um, but what an amazing
1:05:16result. Like if you look at this the
1:05:18detail we get and the the reflectiveness
1:05:20and the you know the the trees and the
1:05:23mountain and the whole Hawkbury right so
1:05:25the context the place that it is that
1:05:28photogometry would always give us that
1:05:30really well but wouldn't give us this
1:05:32wouldn't give us that waterfall or that
1:05:34you know that vista so I've been
1:05:36learning a lot on you know you got to
1:05:38start thinking about shooting everything
1:05:40as well so photoggramometry your mindset
1:05:42is I want to shoot the thing I want to
1:05:44shoot the subject and you forget about
1:05:46the clouds or the trees or the the water
1:05:49around you. You got to shoot that now.
1:05:50And I've been learning hard hard
1:05:52learning, you know, to unlearn some of
1:05:54my things. Water water is hard. Water
1:05:57comes up, you know, in this instance
1:05:58because it was a a windy day didn't come
1:06:01up very well. But generally amazing
1:06:03results and and part of um what we're
1:06:06doing. But if you look at the mesh, the
1:06:08mesh itself, so I did then process into
1:06:10mesh as well and we shot for for mesh
1:06:12and for um splat um photoggramometry
1:06:16still wins on some things. Hard surfaces
1:06:18wins, measurability wins. Um so there's
1:06:21the there's the actual um final mesh
1:06:24that is a photoggramometry model.
1:06:26traditional photoggramometry hard
1:06:28surface model um has come up really well
1:06:30and does a really good job to you know
1:06:33really give you a condition report of
1:06:34that site um which is really what we
1:06:37were trying to do. Someone was living
1:06:38there at one stage. Um but you know the
1:06:40mesh itself doesn't do trees very well.
1:06:42They look blobby. We've all seen blobby
1:06:44photoggramometry trees. The you know my
1:06:46my plan now and what I'm really doing
1:06:48commercially is merging the two. Taking
1:06:50the good bit of the mesh and the and the
1:06:52good bit of the splats and and creating
1:06:54something out of it. But yeah, the
1:06:55bottom right from a boat. Yeah, return
1:06:58to home doesn't work when you're on a
1:07:00boat. That was a scary day trying to get
1:07:02it to land on a boat. Um, which you
1:07:05know, my drone came home with me. I'll
1:07:07put it that way. It was scary but um
1:07:09survivable. And then this one, these are
1:07:11the last of the case studies. These this
1:07:12is really and I've been doing this in
1:07:14parallel the whole time. So this is part
1:07:16of a project coming out in a month or
1:07:17two for Acher, which is the um
1:07:20Aboriginal Culture Heritage and Arts
1:07:22Association of New South Wales. that's
1:07:23an umbrella or of all the cultural
1:07:26centers in New South Wales. And I've
1:07:27been working with them um through a
1:07:29partnership sort of model um on a
1:07:31project to create a virtual cultural
1:07:33center. Um and the first one we did was
1:07:35out at um the Armadale Culture Cultural
1:07:38Center in Keeping Place um Act KP. And
1:07:42this was my first attempt at just trying
1:07:44to do a splat. And I went there early in
1:07:46the morning and I walked around with my
1:07:48camera um taking photos like this. Like
1:07:51I walked and I walked back and I walked
1:07:54back and I walked back. I walked around
1:07:55in circles and then I got my drone out
1:07:57and I flew my drone kind of where I'd
1:07:59walked and I flew it a bit higher. So I
1:08:01flew the drone up a bit and over this
1:08:03way a bit and down that way a bit and I
1:08:06aligned it all and I got this. It's just
1:08:09I I was tears came to my eyes the first
1:08:12time I saw it. I'm like I could you
1:08:14could do this now. I could never do
1:08:16this. Photoggramometry forget it. Like
1:08:18just absolutely not. You could do a 360
1:08:21like Callum said, you know, you could do
1:08:23a place where you could look at this
1:08:24thing place where it was, but being able
1:08:26to walk around and do what I've just
1:08:28done, no, not possible. So, this one, I
1:08:31mean, what what did I forget? I forgot
1:08:33to shoot things like the sky. So,
1:08:34there's all these gaps in the sky I sort
1:08:36of faked afterwards. Um, but it was a
1:08:40pretty good first attempt and it was a
1:08:41pretty good lesson um from where I was.
1:08:44Um, so then my next one was early late
1:08:46January. I went up to um Yarawara um
1:08:49which is uh near Kindi Beach um which is
1:08:53the um keeping place up there um and
1:08:55cultural center. Amazing place. And I
1:08:58shot this down at Pipe Clay Lake. So
1:08:59this is down on the ocean. This is my
1:09:01test footage. And this one came up
1:09:03actually quite well, but I shot it with
1:09:06a drone. I was silly. I was I was I had
1:09:08bravado this morning. So I set my drone
1:09:09to do an automatic, you know, zigzag
1:09:13grid as you do with a drone. And at the
1:09:14same time, I thought I'll shoot it with
1:09:16my Insta 360 into my ground camera all
1:09:18at the same time. So, I was walking
1:09:20around as my drone was flying and I was
1:09:22there. So, I didn't some things I didn't
1:09:24get like I didn't get the roof very
1:09:26well. Um the drone shots were too
1:09:28disconnected from the ground shots. I
1:09:30broke my number one rule of bridging
1:09:33that I had my drones up there in the air
1:09:34and I had my ground down here and trying
1:09:36to, you know, align those two was a
1:09:38problem. But it did this incredible job
1:09:40of the whole place. Like I can now go
1:09:43up. like having the drone shots. I got
1:09:46bits here that um like I didn't even go
1:09:48near that bridge yet. There's the
1:09:50bridge, right? My one of my favorites is
1:09:52over over here in the trees. Like down
1:09:55here you can, you know, go down into
1:09:57this now and see this palm tree. I never
1:10:00went in there. Like that's all from the
1:10:01drone shots flying zigzag like you know.
1:10:03So incredible ability to stay true to
1:10:07photos. Um and and that was sort of my
1:10:09second one. My third one was down at
1:10:11Denny's. So I went down to Deneloquin
1:10:13and this one I got there in the dark. I
1:10:14waited in my car. A few people were
1:10:16camping. I had to wait for them to pack
1:10:18up and go and I shot this
1:10:24which
1:10:26is a new game. It's a river. You got the
1:10:29reflections of the trees.
1:10:33Even if I could make that water plane in
1:10:34photogometry, you're not doing that. You
1:10:36know, you're absolutely not doing that.
1:10:38the trees, the light, the the whole
1:10:41environment, you know, this one worked.
1:10:42And I learned you kind of have to shoot
1:10:45like a crazy person. Like I had to walk
1:10:47around taking shots of everything. Like
1:10:50I I started walking taking shots in a
1:10:52spiral as I walked, thinking I got to
1:10:54take a shot of that and that and that.
1:10:55And I it was a really interesting
1:10:57morning for me. But I I did it
1:10:59structurally. I made sure I was
1:11:00following the photogometry rules so they
1:11:02would align. But no, I needed shots
1:11:04straight up to get all the trees. I
1:11:05needed shots of everything from every
1:11:07angle. Yet at the same time, I needed to
1:11:09be efficient and I needed to do it
1:11:11quickly because the sun was moving. So,
1:11:13yeah, what a wild it was a wild, you
1:11:16know, experiment, but it's an experiment
1:11:18that worked. And it's one that now I
1:11:20sort of have confidence on how to shoot
1:11:22and stuff and how to how to mix the
1:11:23cameras. And then the final one was um
1:11:25up at Goen. So, Goen Water Hole Trangi
1:11:29um in sort of central New South Wales. I
1:11:32shot this
1:11:41So that's not a video. That's, you know,
1:11:44a full 3D resolved place. I shot it in
1:11:4810 minutes.
1:11:50So I shot that with one of the new 360
1:11:52drones. It's a 10-minute flight. The
1:11:55flight itself
1:11:57is here. That's my flight. So I flew
1:12:00slowly. I flew around. I flew through
1:12:03things. I got higher and higher and
1:12:04higher and then I landed and I had to
1:12:07shoot it quickly. Look at the sun.
1:12:09So this this is a new game now. So
1:12:12here's the here's the actual input. So
1:12:15yes, it's lower res. Callum's right. You
1:12:17know, yes, it doesn't resolve details as
1:12:19well, but you know, with enough
1:12:21coverage, you can very quickly get an
1:12:23amazing site. You know, the the lighting
1:12:26on this one is incredible. It's actually
1:12:28not the one I end up using because the
1:12:30one I end up using um that I'll show you
1:12:33in one minute. I'm really getting close
1:12:35to time. I'm doing okay. Um
1:12:39is this one which I shot the same way as
1:12:41the Deniloquin boat ramp where I shot it
1:12:43walking around. I shot it with my normal
1:12:45drone. Um and I got a lot more crispness
1:12:48on the trees. So, you know, and you
1:12:50know, again, Callum's right. you know,
1:12:52shooting with a, you know, good
1:12:53mirrorless with a good lens is is triple
1:12:5710 times what you would get from a 360.
1:12:59So, you know, you'll watch a lot of
1:13:00stuff on the web saying, "Yeah, walk
1:13:02around with a 360 and bang, look at the
1:13:04amazing splat." Didn't hasn't worked out
1:13:06that way for me. But the flying drone
1:13:08has changed my game a fair bit. So, then
1:13:11um but then what do you do with it? So,
1:13:14this is the final couple that I've been
1:13:15writing um utility viewers for this now.
1:13:19So, I'll show you this one first. This
1:13:20is the one I started with. So this is
1:13:23all of a sudden I've got the bottom. No,
1:13:25no, I have the bottom. So this is my
1:13:28experimental I've written experimental
1:13:29viewer that does the best of both
1:13:31worlds. That is a mesh. Right? So that's
1:13:35a photoggramometry model and that is a
1:13:37splat shown in one viewer. So if we do
1:13:40this right, you can merge the two.
1:13:42Right? So you can use mesh for what mesh
1:13:45is good for, splats for what splats is
1:13:46good for, and then put them together
1:13:48into our deliverable. that is this. So,
1:13:50it's measurable or the the pot is at
1:13:52least. Um, it's got a real um solidness
1:13:55to it, yet the feathers are still
1:13:57rendered. This is my dream. This is
1:13:59where I'm pushing. And this is really
1:14:00where I'm putting a lot of R&D at the
1:14:02moment into doing this kind of stuff.
1:14:04And then if you look at what I'm doing
1:14:06in sites for that and in VR, this is um
1:14:09a preview only. This is not done. Um, so
1:14:12I'm showing some preview work. Um, so
1:14:14this is the the virtual gallery that I
1:14:16built down at Denny Boat Ramp. So, I've
1:14:18got a lot of relics that you can pick up
1:14:20in VR and have a look at. Um, but then
1:14:23you can walk out of this into the place,
1:14:26right? And you can look back and you've
1:14:27got this little TARDIS kind of approach.
1:14:29But the place itself, the foreground is
1:14:33a mesh, right? That tree is a mesh. That
1:14:37tree over there is a splat. So, I'm
1:14:39trying to work out ways where you can
1:14:41get the solidness where it's important
1:14:43in the foreground, but the splats where
1:14:45it's important, which is on the on the
1:14:46trees. And if you look closely at some
1:14:48of them, you can sort of see where one
1:14:49stops and the other ends. I've done a
1:14:51pretty good job generally. But, you
1:14:53know, if you go right over here to the
1:14:55edge, you'll see mesh mesh mesh mesh
1:14:57mesh splat. It's really obvious when I
1:15:00show you. And there's a bit of an
1:15:01overlap zone here. So, this is really
1:15:03what I'm pushing now as a as you know,
1:15:06sort of what we can do with this once
1:15:08we've shot it. And it's important now to
1:15:10shoot both ways. So, you shoot for for
1:15:12um polies and you shoot for splats as
1:15:14well. And that's I think certainly in
1:15:16our in our game that's what's important.
1:15:19So that's my showcase. I'm going to
1:15:21quickly run through this next bit and
1:15:22we'll have a quick moment for
1:15:23discussion.
1:15:25You're not going to learn how to do this
1:15:27tonight. Okay? I don't think you should.
1:15:28If anything, take away some of those
1:15:30lessons I've been talking about some of
1:15:31the hard one stuff that I've got. Um but
1:15:34if you do photoggramometry already, you
1:15:36know 90% of it. There's the good news,
1:15:38right? So a lot of it's just the same.
1:15:40We we use the same equipment. We shoot
1:15:42with the same kind of discipline,
1:15:43coverage, overlap. We align things, um,
1:15:47structure from motion, use mostly the
1:15:49same software. So, your habits, your
1:15:51your equipment, all the same. It's good.
1:15:53What's different is, you know, we need
1:15:56more view diversity. Like I'm saying, I
1:15:57start to shoot in a different way or I
1:15:59start to shoot more things than I would
1:16:00have shot normally. It's more tolerant
1:16:02of mixed conditions, but less tolerant
1:16:04of ambiguity. And it's the real
1:16:05conundrum that I've been that I've
1:16:07really learned the hard way, but I've
1:16:08come to love. Like, um, the training
1:16:12after alignment is quite different and
1:16:14you know there's different viewers in
1:16:15delivery at this stage you know we're
1:16:17trying to write it in um but when it
1:16:19comes down to it I'll go back into
1:16:22I go back into slideshow mode
1:16:27um yeah we shoot we align and then you
1:16:28have a choice to go to splat or mesh
1:16:30that's my suggested flow so you know
1:16:32you're following photoggramometric
1:16:34discipline at your capture you're adding
1:16:35view diversity and then you choose both
1:16:38or either or you know that's what you
1:16:40can use it
1:16:41So, shooting um standard cameras are the
1:16:44best, DSLRs, mirrorless still the best.
1:16:47360 lot of options coming out um that
1:16:50you can, you know, handle that, but
1:16:52you're you're in this odd conundrum
1:16:54where it's got to be undistorted for
1:16:55training, but you know, you end up with
1:16:57a 360 distorted image. You have to
1:16:59manage that. Drone works amazing. Um you
1:17:02know, absolutely suited to it. Um a lot
1:17:05of laser assist is coming out as well.
1:17:07So handheld slam, terrestrial lidar um
1:17:10can help and you know most splatting
1:17:13works with a point cloud to start with.
1:17:15Um phones like everything I just if you
1:17:18have to people go can I do it on my
1:17:21phone? Of course you can do anything on
1:17:22your phone. Do we as a last resort? You
1:17:24know if you're in the field if you're in
1:17:26an archaeology dig and you're the best
1:17:28thing you've got is your phone. There's
1:17:30a broader problem on data collection.
1:17:32I'm sorry, but that's just my honest
1:17:34opinion that, you know, like you can
1:17:36with a phone, but as archaeologists and
1:17:38as museum professionals, we don't. Okay.
1:17:41Um, and you can mix them, but like
1:17:42photogometry, you need to test that
1:17:44workflow and and work it out. You've
1:17:45seen ones where I've mixed, you know,
1:17:47drone and and ground cameras, 360 and
1:17:49other cameras.
1:17:51Um, alignment, if you're doing it
1:17:53already, keep doing what you're doing.
1:17:55If you're doing metape, use that. If
1:17:56you're doing reality capture, use that.
1:17:59You don't have to change that. you know,
1:18:01most of what I've shown you today, I use
1:18:03Metashape or Reality Capture for um you
1:18:06know, they're they're great for what
1:18:08they are, but there's a couple open-
1:18:10source players coming into the market.
1:18:12So, co not market, they're free, but
1:18:15Cole Maps one and Glow Maps the other.
1:18:16Cole Maps probably dominant and Cole
1:18:18Maps integrated to a lot of the Gausian
1:18:20splatting software. Um it's slow, but
1:18:24it's pretty good. It does a beat meta
1:18:27shape and reality capture. Not in my
1:18:28mind, but it's pretty good. So, you
1:18:30know, you can get a result with that.
1:18:32Um, but yeah, there's equangular images
1:18:36are problematic the way you split them.
1:18:38Uh, mixed focal lengths can actually
1:18:40cost your alignment quality and
1:18:42alignment is everything. Like alignment
1:18:44is so much more important here. And
1:18:46again, photoggramometry is forgiving.
1:18:48Splatting is not. It stays true. So if
1:18:50you have a slight misalignment, the end
1:18:52result stays true to that slight
1:18:54misalignment, which means a shadow or an
1:18:56echo or things like that. Um, and yeah,
1:18:59you need a sparse point cloud and that
1:19:02that's a good area of development to
1:19:03look at, you know, how that changes
1:19:05things. Training options, few on the
1:19:08desktop, most of them are GPU bound. Um,
1:19:11postshot it used to be free, it's now
1:19:13paid. It's the one I use. It's probably
1:19:14the most friendly if you want a piece of
1:19:16software that is normal for you. Um, if
1:19:19you like scripts and running software
1:19:21yourself and compiling stuff, you got
1:19:23Brush in the middle that's easy to
1:19:25download and compile. It's open source.
1:19:27It works on Mac as well, which the
1:19:29others don't, but there's not it's not
1:19:30quite as friendly. And then there's
1:19:32Lickfeld Studio that's it's newer. It's
1:19:34hardcore. The speed of development is
1:19:36crazy on that. Um, but it's very
1:19:39difficult to install. It's fully
1:19:41featured, which means you need to know
1:19:42how to tweak all the different bits of
1:19:44it. So, if that first definition scared
1:19:46you, that will scare you more. Um,
1:19:49there's a bunch of cloud platforms that
1:19:52go have at it like learn the process,
1:19:54but as with everything, we don't really
1:19:56do that in archaeology because we don't
1:19:59want the data to get out there. We don't
1:20:01really do that in museums because, you
1:20:03know, that it's just not the way we
1:20:05work. So, I mean, they're good to learn.
1:20:07There's a few closed systems like XG
1:20:09grids um which is a a combination
1:20:12hardware software pipeline um offering
1:20:15where you buy a device that's a slam
1:20:17device with cameras in it that links to
1:20:19their own software that links to their
1:20:20own publishing platforms kind of like
1:20:22Matterport I guess you know if you know
1:20:24what Mapport is for real estate um and
1:20:27then there's emerging ones as well like
1:20:283D Zeph's already released a beta it's a
1:20:31matter of time before Metashape and
1:20:32Reality Scam release their own training
1:20:34in in software I think that's what's
1:20:36going to happen in 12 months it's
1:20:37converging. So the tools are going to
1:20:39end up being the tools you know already.
1:20:40It's just going to be another option
1:20:42within those tools to to do splatting.
1:20:44Um so how it really works, it's pretty
1:20:46simple. Um you start with a cloud. Um
1:20:49that's the seed. It then it basically
1:20:52adds little blurry things wherever the
1:20:53cloud is and works out how well they
1:20:56look to the image. It then cuts the ones
1:20:58that are wrong, adds more where it needs
1:20:59more, and repeats that thousands of
1:21:02times. So it's an iterative training
1:21:04process. I'm going to try and be brave,
1:21:06which you know, those that know me know
1:21:09I I can try and be brave. Let's just
1:21:12let's get on training right now. So,
1:21:14I'll open post shot. There's a post shot
1:21:17there. So, this is what post looks like.
1:21:19It's got a gooey. It is, you know,
1:21:21fairly friendly. I can get um some
1:21:24photos. So, here's a stack that I've
1:21:26done before. It's a set of images of the
1:21:29puppets. It's a sparse cloud and it's a
1:21:32definition. So, I can drag those onto
1:21:33here.
1:21:34Um, I gave it a bunch of options. I
1:21:36won't start training yet. So, if I
1:21:38import that, I'll just quickly fix the
1:21:41rotations. Come in. Come in Zed up, not
1:21:43Y up, which is what it does. Is that
1:21:46going to stop me?
1:21:48Oh, num locks off. That'll stop me.
1:21:52So, you can see here I've got cameras,
1:21:54I've got points. This is should be
1:21:55normal if you're in photogometry. I'll
1:21:57turn my cameras off and show the points.
1:21:59So, that's what seeds it. And then from
1:22:01those, it'll throw gorians there in that
1:22:03color. and then iterate from there. So,
1:22:05let's start it training. It's going
1:22:07good. So, you can see instantly it's
1:22:09throwing a bunch of splats everywhere.
1:22:11And if I turn the points off, it starts
1:22:14quite blurry and that will converge on
1:22:16an accurate representation of that from
1:22:18every angle. And that will converge on
1:22:20those wires and the fuzziness and all
1:22:23the rest. And it will probably, you
1:22:24know, I usually train for 100,000 steps.
1:22:26It might take, you know, um
1:22:30half an hour, an hour, but the the main
1:22:32bit is um look at my graphics card. It's
1:22:35gone to 100%. This is what it does. The
1:22:37better graphics card you have, the more
1:22:39graphics memory you have, um the better
1:22:41training speed you're going to get. And
1:22:43that's the way it works. So, that's a
1:22:45quick initiation that did not crash.
1:22:48Look at that. Winning. Um
1:22:51okay, so clean up. Cleanup's a problem.
1:22:54There's no easy way to clean them. Um,
1:22:56you can't decimate them. That doesn't
1:22:58work. They're view dependent. So, yeah,
1:23:01we can get rid of floaters. We can do
1:23:02some stuff with it, but not a lot. So,
1:23:05that's, you know, you know, coming and
1:23:07with no decimation. How do we deal with
1:23:09levels of detail? How do we deal with,
1:23:11you know, the deliverable versions? So,
1:23:14you know, there's a bunch of formats
1:23:15that are standardizing as we speak. Um,
1:23:19some have now been added to glTF, but
1:23:21again, that itself is a bit of a
1:23:22minefield. So, what kind of formats are
1:23:25they? But basically, like the line at
1:23:27the bottom, whatever you publish today,
1:23:28you're going to need to do it again in
1:23:30the future. It's going to change in the
1:23:31next 12 to 18 months. Um, and then
1:23:34publishing, Super Splat's the main one
1:23:36that people use. It's online. Um, Play
1:23:38Canvas is what Super Splat was built on.
1:23:41It's a a game engine, a web based game
1:23:44engine. It's what I'm using to write my
1:23:46own viewer. Um, and it's what Superlat's
1:23:49written on. You can use 3JS with a Spark
1:23:51integration. those ones I showed you of
1:23:53Denny with the the virtual cultural
1:23:55center in it. That was 3JS with Spark.
1:23:58Um, and we're adding it to Pedestal
1:23:59right now. So, later this year, Pedestal
1:24:013D will have it in and you'll be able to
1:24:03do splats and meshes and all of that
1:24:05that way. A bunch of game engines it
1:24:07goes into as well. Um, but mobile
1:24:09performance is a problem. It's, you
1:24:11know, it's still quite computationally
1:24:13heavy to render them in in screen. All
1:24:15mine look nice tonight, but I've got big
1:24:17graphics cards. So, you know, how do we
1:24:19get them to people without that? Um,
1:24:23so I'll just I'll I'll cover over this
1:24:24because I want a bit of time for
1:24:27discussion. So, um, keep your legacy
1:24:30data, play with it, use it. Um,
1:24:32alignment is everything. If you're not
1:24:34doing it now, consider running better
1:24:36optimizations than you're currently
1:24:37doing. I I sort of went straight into
1:24:39the um the the Geology Society, USGS
1:24:44approach to alignment correction in
1:24:46photoggramometry that helped some of my
1:24:48results on the boats. Um, but
1:24:50experiment, iterate, experiment. You
1:24:51know, it's just like photogometry was
1:24:53before this. You got to you got to get
1:24:55you got to get tacit. Um, you know,
1:24:58there's all kinds of stuff about, you
1:24:59know, what it is and what it does for
1:25:02us. But, you know, the big one is the
1:25:04one at the bottom. It's not replacing
1:25:05photoggramometry. It is joining it. It
1:25:07is just something else that we can do
1:25:09now. So, finally, be like Bob, right?
1:25:13Learn to love the happy little splats.
1:25:14learned to love the the blurriness and
1:25:17the you know the the weirdness. Thank
1:25:19you Callum. Of of what we're doing here.
1:25:22Um you know you you get past that and to
1:25:24me it's transformative. It's changed the
1:25:26way I look. It's changed the way I see
1:25:29um as a photoggramometrist. It's really
1:25:31it's liberating. So let's open for
1:25:34discussion.

About this transcript

This page contains the full transcript of Demystifying Gaussian Splatting for Archaeology and Museums | CAA Australasia Seminar | 18 June 2026 by CAA Australasia, generated from the public captions YouTube serves with the video. The transcript has 16,977 words across 2,325 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.