Digitally Reconstructing Mawson’s Huts, Antarctica.
With the ongoing improvement in AI and machine learning models, as well as adaptive, programmable workflows in environments like Python/Jupyter, ComfyUI, Blender and Unreal Engine, it is a productive time to look at new approaches in heritage visualisation and digital conservation.
Over the last 20+ years I have accumulated an extensive database of images of the Cape Denison heritage site, including extensive external and internal imagery of Mawson’s Huts.
This database includes not only my own photographic, video and audio documentation, but also the work of many other expeditioners to the site, courtesy of the first-rate people of the Mawson’s Huts Foundation, Australian Antarctic Division and IPEV. I am especially grateful to Rob Easther, who facilitated my two trips to the site, and also enabled me to work with members of other teams to generate structured image sets suitable for photogrammetry. Rob was a wonderful man who I admired greatly and counted as a friend – sadly, he passed away in February 2024. Deservingly, he actually has an island in Antarctica (humorously) named after him – Easther Island. I also want to acknowledge Bruce Hull (formerly of the AAD) for employing me to document the site in 2006/7 and providing access to many documentary materials. This led to my first work for fulldome: ‘Frozen in Time’ (2008).
As a consequence of these opportunities and connections, I have a database containing imagery from about 2002 to 2021. These are largely unordered collections – though with enough coverage to enable the assembly of multi-year sets that can be used for partial 3D reconstruction. Photogrammetry typically doesn’t work well with non-contiguous data (both spatial and temporal), but with masking it is possible to work around many registration issues. It’s a bit of a kludge, but seems to be working. The creation of masks is typically very labour-intensive, but fortunately I am now able to semi-automate the processing of thousands of files using AI text-vision models like SAM/SAM2 and Florence-2. Semantic segmentation and masking can also be semi-automated using locally installed private LLM systems like Ollama/llama3.2. ‘Semi-automated’ means ‘fragile’ – still requiring the human-in-the-loop, in concert with a vector database for embeddings in latent space. Combining these with other opensource processors like COLMAP, SuperSplat, NerfStudio, one can imagine a future ‘agentive’ workflow for dealing with large unordered image collections – similar to the approach griptape.ai is developing, but without the reliance upon external paid APIs. Interesting times indeed!
Over time, I’ll post a few works-in-progress here, including photogrammetric and gaussian splat reconstructions, which will ultimately lead to a full 3D reconstruction in Unreal Engine. It’s been a long time coming, but the objective I’ve held for many years is now within sight. Credits for photographic sources are listed at the bottom of the page.
The quality of reconstruction varies due to the quality and source of the underpinning materials – for instance, some image sets from 2002 were shot on a digital camera at 1600 x 1200 px resolution; later ones at 5472 x 3638 px – so the resolution is quite variable. This broadly matches the development of digital camera technology over the last 20+ years. Interesting possibilities arise with ‘up-rezzing’ source images using generative AI models, at the risk of introducing confabulated detail. It’s a trade-off between digital ‘accuracy’ and novel possibilities for generative reconstruction. Of course, I am aided by having personal experience of working at the site and witnessing its changes over many years.
Similarly, as the conditions at the site are very variable, creating a representative 3D model is quite challenging. For instance, photographs taken in 2006 (Simon Mossman) reveal the Huts to be almost entirely entombed in ice (yes, ice, not snow), whereas conditions in 2021 (Dave Killick) revealed the structure to be almost standing free from the surrounding rocks.
Similarly, the interior of the main hut has had massive amounts of ice removed over the years – thanks to ongoing conservation expeditions by the Mawsons Huts Foundation – transforming its appearance – and revealing many of the treasures previously entombed within. I hope to eventually be able to create a gradual ‘reveal’ demonstrating the amazing conservation work that has been undertaken by the talented archaeologists, materials conservators and heritage carpenters who have worked there.
Of course, it would be wonderful to be able to return to the site myself with a very high quality full-frame or medium format camera and shoot everything at 100-200 megapixels, or in lieu of that, to have team capacity to do the job in a technically correct way.
In addition to my own extensive materials (2006/7 and 2009/10), I’m very grateful to for the photographic work of Adrian Welke (2002), Simon Mossman (2006), Chris Henderson (2009/10), Sally Hildred (2016), and Dave Killick (2021). I’m always on the lookout for more – so if you’re reading this and have some good photographic documentation – please drop me a line. Hopefully, in future expeditions we can arrange a thorough photogrammetric survey. In the meantime, I’ll keep plugging away with what I have, figuring out new ways of doing things with it – that’s half the fun.
