Every capture on this page starts the same way: a camera moves around a place and takes hundreds of overlapping pictures. What happens next decides whether you get a model you can measure, a scene that looks photographically real, or both.
Photogrammetry: geometry you can measure
Photogrammetry is the oldest of the three methods. Software finds the same features in overlapping photographs, works out where each camera was, and triangulates the features into 3D points. This is usually called structure-from-motion followed by multi-view stereo. The points become a dense point cloud, then a textured mesh. Viewed straight down, they become an orthomosaic: a single stitched map with consistent scale.
Bart's own definition is that photogrammetry "builds measurable 3D models from overlapping photos, often captured by drone"12. Because the result is real geometry, it is the format that engineers, GIS analysts and game engines expect. In his Riyadh post he puts the trade-off plainly: a photogrammetry mesh "is better for measurement and for use in game engines and GIS tools"11.
Photogrammetry struggles wherever surfaces do not behave like matte paint: glass towers, water, polished stone, thin railings, foliage. Features that change appearance with the viewing angle cannot be matched reliably between photos, so those areas tend to come out melted, smeared or full of holes.
NeRF: a scene learned by a neural network
Neural radiance fields, introduced by Mildenhall and colleagues in 202031, take the same posed photographs and train a small neural network to predict the colour and density of any point in space, as seen from any direction. You can then render the scene from viewpoints no camera ever occupied, with reflections and fine detail that photogrammetry misses.
The cost was speed. Early NeRFs took a long time to train and were slow to render, which made them awkward to publish on the web. Bart's 2023 Riyadh notes describe drone, NeRF and photogrammetry work side by side12, and his LinkedIn from that period includes NeRF posts37.
3D gaussian splatting: photoreal and real-time
3D gaussian splatting, published by Kerbl and colleagues in 202332, keeps NeRF's photographic quality but replaces the neural network with millions of small, coloured, semi-transparent ellipsoids: the "splats". They are optimised against the source photos and then drawn directly by the graphics card, so a scene renders in real time, even in a web browser or on a phone.
Bart's use-case post sums up the appeal: splatting "captures reflective, thin and organic detail that photogrammetry misses, and renders in real time on ordinary devices"7. Splats are usually stored as .ply files, the same container used for point clouds, and the same post mentions converting .ply to lighter .splat files by drag-and-drop for web viewing7.
The trade-off is the reverse of photogrammetry's. A splat looks right but is not a clean surface that can be measured, simplified or 3D-printed. About his Riyadh captures he writes that splats "have decreased quality but offer extra features: like light and reflection preserveration at night"11. That property shows clearly in his many night-time captures of London3.
Where the drone fits in
All three methods need the same raw material: sharp, well-exposed photographs from many angles, with plenty of overlap between them. A drone makes that cheap at the scale of a building or a neighbourhood. A grid flown with the camera pointing straight down gives the even, top-down coverage an orthomosaic needs. Orbits and angled ("oblique") passes add the facades and vertical detail that meshes and splats need.
The reconstruction does not care what carried the camera. The Burgess Park models were made from "drone or phone footage"4. In his marine-biology post the camera is held by divers, who photograph coral along orbital swim paths around each colony8.