Calibrated Per-Carrier Confidence and Certified Pruning for Gaussian-Splat Assets

Abstract

A 3D Gaussian-splatting asset ships millions of primitives, each carrying a per-primitive confidence — a view count, an opacity — that downstream pruning, streaming, and level-of-detail decisions consume as if it meant something. It does not: these numbers are heuristics with no operational meaning and no guarantee, so a consumer that drops “low-confidence” primitives is guessing. We ask what it takes to give a splat asset a per-carrier confidence a consumer can actually trust, and report AURA, a preview-stage library (v1.0.0) that attaches one. The method has three parts, each pure-CPU post-processing over the exported carrier array. First, a held-out per-carrier reliability label from multi-view re-projection agreement: a carrier is reliable to the degree its stored colour matches what independent held-out views observe where it projects, computed under both an occlusion-blind colour label and a stricter occlusion-aware depth label. Second, isotonic (PAVA) calibration of an export-time feature — train-view colour agreement, computable without held-out ground truth — into a calibrated confidence. Third, a distribution-free split-conformal pruning certificate: prune every carrier below a threshold (\tau) and the retained set’s mean unreliability is at most (\epsilon) with confidence (1-\alpha).

On four real scenes (Tanks-and-Temples Truck; Mip-NeRF-360 Garden, Kitchen, Room), the export-time feature predicts held-out reliability at correlation 0.91–0.98 (colour label) while the shipped view-count heuristic ((|r| \leq 0.13)) and opacity (−0.18 to +0.16) do not; isotonic calibration cuts expected calibration error from 0.46–0.59 to 0.0006–0.0016 (~300–900×); and confidence-guided selection recovers 0.58–0.72 mean retained reliability, within 1–4% of an oracle ceiling and above opacity (0.37–0.53, at or below random) at every budget — keeping only 10% of carriers it retains 0.77–0.90 reliability versus opacity’s 0.31–0.49. The properties survive the stricter occlusion-aware label; they further persist at native full resolution under a corrected, genuinely held-out split (which lowers one over-optimistic per-scene certificate from 1.00 to 0.77), transfer across scenes for selection (AUC within ±0.0004) with a small per-scene conformal set restoring the certificate, and survive a stricter render-loss label read from the actual alpha composite with honestly weaker margins. We extend the single pruning certificate into a certified level-of-detail ladder — a consumer streams carriers in descending calibrated confidence and may stop at any published level with a distribution-free, family-wise bound on the unreliability mass it thereby discards (all 16 bounds hold across four scenes and four levels). We deliver the calibrated confidence through the KHR_gaussian_splatting_AURA_CONFIDENCE channel, the OpenUSD 26.03 UsdVolParticleField3DGaussianSplat schema, and a sidecar for Niantic’s SPZ v4 container whose writer is cross-validated bit-exact against the reference implementation. We are deliberate about scope: AURA is a preview-stage system; the reliability labels are self-computed re-projection proxies, not human ground truth; the evaluation gates, while content-checked against the committed artifacts, are our own and run locally; the scenes are single captures; the typed-carrier backend result we report (+0.33 dB on Truck) is a reproduction of Deformable Beta Splatting, not a contribution of this paper. What we do claim is narrow and, we believe, new: a calibrated, certified, exportable per-carrier confidence is a capability a bare Gaussian splat does not have, and this paper measures it honestly on real scenes.