The Neural Turn in Capture

To see what is new, hold the old methods clearly in mind. Laser scanning measures distance directly: fire a pulse, time its return, record a point. Photogrammetry measures indirectly but still geometrically: find the same feature in many overlapping photos and triangulate where it must be in space. Both end in the same currency — explicit 3D geometry, points and surfaces with coordinates you can measure against. The computer stores *where things are*.

Neural capture stores something different: it stores *how the scene looks*. You feed a machine-learning model many photographs of a subject, each tagged with the camera position it was taken from, and the model trains until it can reproduce those photos. In doing so it learns an internal representation of the scene's colour and structure rich enough to render the scene from a brand-new camera position it was never shown. This is novel view synthesis — inventing a plausible photograph from a viewpoint that never existed. The two landmark techniques are NeRF (2020), which stores the scene inside the weights of a small neural network, and Gaussian splatting (2023), which stores it as millions of little coloured 3D blobs; Modules 4.2 and 4.3 take each in turn.

The shift is profound. Classical capture answers "where is this surface?" with a measured coordinate. Neural capture answers "what would a camera see from here?" with a rendered image. The first is a ruler; the second is a painter with an extraordinary memory. Both are built from the same raw ingredient — photographs or scans of the real world — and both belong to the broad field of computer vision and 3D reconstruction. But their outputs are not interchangeable. A point cloud is a measurement you can audit; a neural render is an appearance you must take on trust until you verify it. Understanding that one-sentence difference is the whole foundation of this module, and the reason a beautiful neural capture is never, by itself, a survey.

The excitement is real and worth naming precisely, because neural capture is not hype for its own sake — it does things the older methods genuinely struggle with. The headline is photorealism. A laser scan of a carved temple column gives you exact geometry but a rather clinical, sometimes patchy appearance; the colour is sampled coarsely and the subtle way light plays across the stone is lost. A neural capture of the same column can look like a film of it — every highlight, every soft shadow, the sheen of polished surfaces, all reproduced so faithfully that a viewer forgets they are looking at a reconstruction.

The second gift is view-dependent appearance, the thing that makes surfaces look alive. Real materials do not look the same from every angle: a glazed tile, a mirror, a lacquered door, water, a glass facade all change as you move. Classical geometry-plus-texture handles this badly, because a single colour is baked onto each point. A NeRF learns colour *as a function of viewing direction*, so reflections and glints move correctly as you orbit the scene — something point clouds and meshes simply cannot do. It also means neural methods cope gracefully with the soft, fuzzy, semi-transparent things that wreck laser scans and photogrammetry: foliage, hair, lace screens, fine grilles, thin edges.

Third, it is accessible. The raw input is just photographs or ordinary video — no lakhs-rupee scanner, often just a phone. For visualisation, presentation, virtual and augmented reality, marketing walkthroughs, and capturing the *visual richness* of a space that a point cloud flattens, neural capture is extraordinary: you can hand a client an immersive, fly-anywhere, photoreal model of their building captured in an afternoon. That is a capability the profession did not have five years ago, and it is why the field has moved so fast. Hold the excitement — and keep reading, because the same properties that make it dazzling are exactly what make it unreliable for measurement.

Here is the discipline this whole module exists to instil. The very mechanism that makes neural capture beautiful — optimising a model until it *reproduces the photographs* — optimises for appearance, not for metric truth. The model is rewarded for rendering images that match the input photos. It is not rewarded for getting the real-world distance between two walls correct. So a NeRF or a splat can look flawless from every trained angle while its underlying geometry is soft, locally distorted, or simply invented in places the cameras did not see well.

Three specific cautions follow. First, scale is not inherent. Classical capture ties into real units through measured distances or survey control; a raw neural capture from phone photos has no idea whether the room is three metres wide or thirty, unless scale is imposed externally. Second, geometry can be confidently wrong. Neural methods produce "floaters" (wisps of density hanging in empty air), smear thin objects, and guess at hidden or poorly-photographed regions in ways that look plausible but are not measured. A flat wall may render perfectly yet not be flat in the model. Third, there is no honest error figure you can read off. A survey-grade laser scan comes with a stated accuracy; a neural capture comes with a subjective impression of realism, which is not the same thing and can be dangerously persuasive.

The professional consequence is simple and non-negotiable, and it echoes the whole course: treat a neural capture as a visualisation, not a survey. If a dimension off a neural model will drive a cut, a fabrication, a boundary, a setting-out or a structural decision, you do not trust it — you verify that dimension against a laser scan, a photogrammetric model built with control, a tape, or a licensed surveyor. Neural capture is a superb way to *see* a place and a poor way to *measure* it, and the rest of this module (4.2, 4.3, 4.4) unpacks exactly why, method by method, and where it does belong.

It helps to place neural capture on the same map as the methods you already know, rather than treating it as a rival that wins or loses outright. Think of two independent axes. One is photorealism and visual richness — how convincingly the result reproduces the look of the real place. The other is metric reliability — how much you can trust a dimension taken from it. Laser scanning sits high on metric reliability and modest on photorealism: superb for measurement, visually plain. Photogrammetry sits in the useful middle: measurable when done with control, with good texture. Neural capture sits high on photorealism and, for now, low on guaranteed metric reliability: dazzling to look at, not to be dimensioned blindly. No method is "best"; each occupies a different corner, and mature practice chooses by what the job needs — or combines them, as Module 4.4 will show.

There is one more property of this field you must internalise: it moves extraordinarily fast. NeRF is from 2020; Gaussian splatting, which largely eclipsed it for real-time use, is from 2023; the literature churns monthly. Techniques that impose real-world scale, extract clean meshes, remove floaters, and tie neural captures to survey control are active, rapidly improving research. So every specific limitation stated in this module is a *snapshot as of writing*, not a permanent verdict. The honest position is to learn the principles — why appearance and measurement diverge, what each representation is and is not — so that when the tools improve, you can judge the new claims for yourself rather than believing a vendor demo.

Across the next three lessons we go deeper: 4.2 builds an intuition for NeRF — a scene as a continuous learned field of colour and density. 4.3 does the same for Gaussian splatting — the faster, real-time representation that took the field by storm. 4.4 draws the honest conclusion: where neural capture genuinely belongs (visualisation, presentation, VR/AR, view synthesis, hard-to-scan visual richness), where it does not yet belong (anything metric or binding), and how it complements rather than replaces measured capture. Keep the course's through-line in view throughout: capture is measurement, you must always know and check the accuracy, and binding results go to a licensed surveyor.