Capture overview

Every splat starts as a folder of photos, and everything downstream, COLMAP’s camera poses, the Gaussians that get densified, the detail a viewer can actually show, inherits whatever that folder does or doesn’t contain. Training can polish a mediocre capture; it cannot invent coverage that was never taken. This article sets out the mindset that makes the rest of the capture part make sense: think in terms of coverage and reconstructability before you think in terms of camera settings or photo counts.

Coverage first, not photo count

It’s tempting to plan a shoot around a target number, a hundred photos for a small object, three hundred for a room, but the number is a symptom, not the goal. The actual target is that no surface in the scene goes unseen, and no surface is seen from only one angle. A capture with fewer, well-placed photos that hits full, multi-angle coverage beats a larger set that’s dense in the easy areas and thin everywhere else. When you’re planning a shoot, walk the space first and ask what a camera path would need to touch to see every surface at least twice, from different positions, rather than asking how many frames you can fit in.

Overlap: how much is enough

overlapThe fraction of a photo's frame shared with its neighbours in a capture sequence. Structure-from-motion needs each surface point visible from several overlapping viewpoints to triangulate its position, making overlap the primary lever for reliable COLMAP registration. is the fraction of a frame’s content shared with its neighbours, and it’s the single biggest lever on whether COLMAP’s structure-from-motion step succeeds. Capture guidance across the current generation of tools, community 3DGS write-ups, mobile scanning apps, and drone flight-planning references, converges on roughly 70-80% frame-to-frame overlap as the working range for general capture, with detailed or building-scale facade work often pushed to 80% or higher because vertical surfaces and fine detail need more redundancy to register reliably. Sources don’t agree on an exact number, some mobile-scanning guidance is comfortable as low as 50-60% for simple, texture-rich objects, but they agree on the shape of the curve: at around 70% overlap a given feature typically appears in three to five photos, which gives COLMAP enough independent views to triangulate its position with confidence. Push overlap down toward 30% and a feature might appear in only one or two frames, which is often not enough: COLMAP either fails to register the camera that took it or places it with a pose that’s noticeably off, and that error propagates into every Gaussian anchored near it. The practical floor, repeated across nearly every capture guide, is that every surface should be visible in at least three images from meaningfully different viewpoints, not three images taken a few centimetres apart.

Parallax: why you have to move the camera

Overlap alone isn’t sufficient if the camera isn’t actually translating between shots: the fix is parallaxThe apparent shift in a feature's position between two photos caused by the camera's optical centre moving through space. It is the signal structure-from-motion triangulation depends on; rotating a camera on a fixed point produces overlap without usable parallax., the apparent shift in a feature’s position between two photos caused by the camera’s centre moving through space, and it’s the signal COLMAP’s triangulation depends on. Spin on the spot to photograph a room and every frame shares enormous overlap with its neighbours, but if the camera’s optical centre barely moves, there’s almost no parallax to recover depth from, and the geometry that comes out is flat or unstable even though the coverage looked generous. The fix is built into how you move: walk an arc or a path with real lateral displacement between frames, not a pan from a fixed tripod point, so every pair of overlapping photos also carries a usable baseline.

Reading a scene before you shoot it: easy vs hostile

Some scenes reconstruct almost regardless of technique; others fight you at every step, and it’s worth learning to spot the difference before you start shooting rather than after a failed COLMAP run. Hostile scenes share a few traits: reflective or transparent surfaces (glass frontages, mirrors, polished floors) that show a different image depending on viewing angle rather than a consistent one; repetitive, self-similar patterns (identical window bays, a repeated column grid, uniform brick coursing) that give the feature matcher plausible-looking but wrong correspondences; large low-texture expanses (plain plaster, tarmac, clear sky) with nothing distinctive to match at all; and anything that moves between frames, traffic, pedestrians, foliage in wind, changing cloud shadow. Easy scenes are the inverse: varied, static, matte, diffusely lit surfaces with natural texture at every scale. Most real sites are a mix, and the useful skill is spotting which parts of a given site fall into the hostile category so you can plan extra passes or anchor shots there specifically.

In practice

Before a shoot, walk the site and mentally flag the trouble zones: a glass lobby door, a repeated facade bay, a plain rendered end wall. Plan a couple of extra passes or supplementary shots there rather than treating the whole scene as uniform. When a COLMAP run comes back with a chunk of unregistered images or an obviously warped section of geometry, check that section against this list first, low overlap, insufficient parallax, or a hostile surface, before assuming it’s a training problem. Nearly every registration failure traces back to one of these three causes, and none of them are fixable once training starts.