Floaters can arise from pose errors, weak coverage, reflections, moving objects or the training representation. Bundle adjustment optimizes the consistency of available observations; it does not guarantee a physically correct solution when matches or constraints are unreliable.
What bundle adjustment optimizes
Bundle adjustment adjusts camera and point parameters to reduce disagreement between observed features and their projections. It works with established observations; it does not inherently know that an armrest must be a particular physical shape.
Calibration parameters may compensate for other errors. Record lens settings, zoom changes and whether images were already undistorted before deciding which images should share intrinsics.
Use several diagnostic signals
Inspect reprojection residuals together with observation counts, track lengths and camera distribution. Low error in easy regions can coexist with weak support for the surface you care about.
Compare the recovered trajectory with the actual capture route. Jumps, reversals, suspicious scale changes and disconnected models warrant inspection, but none is a universal automatic rejection threshold.
Why fully automatic repair remains difficult
Repair requires deciding which matches and priors to trust and whether an image should remain. Sometimes the problem is a pose; sometimes the required surface was never observed. Moving a camera icon cannot create that missing evidence.
Rematching, better calibration grouping, known poses and external geometry are possible inputs, but they require compatible coordinates and meaningful uncertainty. Existing tools already support parts of this workflow; reliable integration across failure cases is the harder problem.
Make correction experiments auditable
Change one factor at a time and record dataset, software, calibration and held-out views. Inspect SfM before comparing splats with the same training setup.
If only trained views improve while side views remain unstable, revisit coverage and materials rather than endlessly minimizing a global average residual.
Sources and scope
Research synthesis and a proposed diagnostic workflow. Not an unprovided Visuloma case study or benchmark.
Sources checked on 16 October 2026. Capture conditions, methods and versions affect applicability.