The capture fix wasn't the mapper, it was the walk speed

A reconstruction that had been fragmenting into disconnected pieces no matter how the processing software was tuned turned out to have nothing wrong with the software. The video feeding it had too many frames.

Photogrammetry reconstructs a scene by matching the same surface point across frames shot from different angles, then triangulating its position from how that point shifts between them. That triangulation needs the camera to have moved between frames, a real change in viewpoint, not just a new frame number. Video shot at a normal frame rate and processed frame-by-frame doesn't give it that. Consecutive frames a fraction of a second apart are nearly the same photo, motion blur and rolling shutter included. The camera barely moved, so there's almost nothing to triangulate from, and the software ends up trying to build a 3D structure out of a wall of near-duplicates. That's what fragmentation had been diagnosing: not a bad setting, a dataset with no usable spacing between its own frames.

The fix was to resample the footage down to a fixed number of frames before reconstruction, a few hundred instead of several thousand, so consecutive frames are seconds apart instead of a fraction of one, with real camera movement and real parallax between them. Once that's in place, the reconstruction software needs none of the workarounds that had been layered on trying to compensate for the density problem. It just works, faster, on a fraction of the data.

The part worth changing in how footage gets shot: that resampling targets a fixed frame count regardless of how long the source video runs. A ten-second pass and a thirty-second pass of the same object both get resampled down to the same few hundred frames. The extra twenty seconds don't add density, they just get discarded in between. So filming longer doesn't buy a denser, more detailed reconstruction. What it buys is coverage: reaching more of the object or walking further around a room, and nothing past that. The variable that matters is how much ground the camera covers per frame once it's resampled, which means how fast the camera was moving relative to the object, not how many seconds the recording ran.

Going forward: move slowly and deliberately, especially close to the object or on a tight loop around it. Don't add length to a pass hoping it captures "more"; it doesn't, past whatever new ground gets covered. If a piece needs more coverage, that's a case for a slower, wider pass or an additional one, not a longer one at the same speed.