Daejeon, July 27–31
We used to wonder where war lived…
War is an effective producer of two major goods: death and documentation. As historians (and digital humanists), we rely on this documentation to reconstruct, recreate, and reflect upon the atrocities committed in the name of whatever cause de jure. The insights we can draw are determined by two key factors: available source materials and the methodologies with which we can investigate them. Digital Humanities methods and distant reading give us a type of insight that more traditional approaches would not allow, as projects like WarVictimSampo (Rantala et al. 2022) demonstrate. Each methodological development has a commensurate gain in insight. With this in mind, this research takes advantage of recent improvements in digital photogrammetry/videogrammetry – effectively creating 3D models from photos or video – to take a much more literal approach to historical reconstruction and recreation, specifically using wartime photos and videos. It combines these with advances in AI to create a streamlined workflow for war photogrammetry - the first of its kind.
This work expands upon a few different bodies of research. The first is research along the lines of Šafář, Staňková, Pospíšil, and Kaňa, who dive into approaches to geolocate wartime aerial reconnaissance images to create orthomosaics, offering some methodological approaches using these same sources. The second is research on retrophotogrammetry, in particular a number of papers by Francesca Condorelli (Condorelli 2023), which outline some conceptual and concrete best practices, as well as providing one option for a general approach. They also make a compelling case for the value of retrophotogrammetry itself, so could be viewed as the motivation for this whole undertaking. The third is the body of various attempts at guides and guidelines for different approaches to photogrammetry; from the 2011 python library through much more recent tutorials on tools like OpenDroneMap.(Bohm 2023) Fourth, ongoing work in integrating AI/computer vision into photogrammetry and aerial photography work, such as work by Cheng et al. on AI for object detection in aerial photos(Cheng et al. 2024) and by Cosido et al. on computer vision in cultural heritage.(Cosido et al. 2014)
It is clear that work has been done broadly on this topic, but there are nonetheless some clear gaps. The most obvious is thematic - no research on retrophotogrammetry takes advantage of the wealth of documentation that conflicts have produced. The secondary gap is methodological - there are not yet clear guidelines or pipelines intended for easy retrophotogrammetry, and even less so with a focus on the use of free and open source tools. The ease of use and access question is a big one; paid tools like Agisoft Metashape are significantly easier to use than open source alternatives like OpenDroneMapper. There are further gaps on the AI front; while some research has been done into the use of AI to augment outputs in photogrammetry work, there are no approaches that line up with the use cases offered in this research.
Given the availability of footage and images, this research looks at three conflicts: the First Gulf War (1990/1991), Vietnam War (US involvement ~1965-1973), and World War II (we all know). Beyond simple questions of material availability, this selection spans a long enough time to give a spectrum of image quality and sharpness. In effect, this allows for testing the limits of retrophotogrammetry as a viable method going further and further back.
This selection gives the blessing and curse of abundant material - far too much for a single researcher to evaluate and examine. Not all footage works for retrophotogrammetry: important here is that it must be of sufficient quality and there must be camera movement. A rotating panorama from a tripod does not provide enough data, while a camera circulating around a building, or someone taking pictures while walking, or footage from a plane all have potential to work. This is where AI comes into the workflow - as a filtration mechanism to isolate which sets of photograms or segments of footage could be worth further work. Where formerly this would be a human-in-the-loop process of identifying each viable piece of archival material, this allows for prompting to identify footage with camera movement, relatively little noise, and the same object/location, among other attributes. By doing this both iteratively and in parallel on the same material, this can be a valuable approach without running into the typical pitfalls of non-deterministic AI outcomes.
The workflow for this process is then:
Tests of this methodology on other archival footage have proved fruitful, including early 90s urban street footage.
Outputs like the ones this work aims to have allow types of research, both quantitative and qualitative, that have been thus far inaccessible - things like examining the viewshed from the top of a reconstructed structure, or examining the destruction wrought in a bombing run. Equally importantly, these outputs allow for a significantly more tangible and experiential type of interaction with historical documents - bringing old photos and footage to life in a way that was impossible until now. A body of scholarship exists on these uses for photogrammetry, and while they tend to focus on these uses for current and ongoing heritage projects, the same benefits apply here - advancing interest and understanding in novel ways.(Portalés et al. 2009; McCarthy 2014; Xing et al. 2025)
And when all is said and done, we will finally see where war lived.