# Semantic and uncertainty-aware splat editing: a baseline note

## Abstract

Interactive Gaussian-splat editing often combines three distinct problems: geometry selection, appearance preservation, and semantic interpretation. Simam Splat Surgeon isolates the first problem in a CPU-safe scalar-property PLY baseline. It exposes confidence thresholding, spatial clipping, region masks, and deterministic point budgets while recording every choice in an edit manifest. This makes the system useful as a reproducible control condition for future learned semantic or uncertainty-aware editors.

## Problem framing

For a splat set \(S = \{s_i\}\), each element contains a center \(\mu_i\), appearance parameters, opacity, and (in Gaussian formats) scale and rotation. A semantic editor would additionally estimate a label distribution \(p(y_i \mid s_i, I, C)\) from splat features, rendered views \(I\), and optional camera/calibration context \(C\). An uncertainty-aware editor should retain uncertainty rather than collapsing it into a hard mask. A simple selection rule is:

\[
M_i = 1[\max_y p(y_i=y) \ge \tau_s] \, 1[u_i \le \tau_u] \, 1[\mu_i \in B],
\]

where \(B\) is a spatial crop, \(u_i\) is predictive uncertainty, and the thresholds are logged. This Space intentionally does not pretend that opacity is semantic confidence: it uses opacity/confidence only as a transparent heuristic and labels that boundary in the UI and manifest.

## Baseline and proposed comparisons

The current baseline applies, in order: confidence threshold → crop → optional sphere/box mask → deterministic budget selection. The budget selector keeps high-confidence points and samples the remaining points across original ordering to avoid a single dense cluster dominating the export. Future systems can compare against:

1. random thinning at the same point count;
2. opacity-only pruning;
3. view-aware importance sampling;
4. learned semantic masks with calibrated uncertainty;
5. uncertainty-aware abstention, where ambiguous points are exported to a review set rather than deleted.

## Evaluation protocol

Report point count, retained volume, and edit latency, but do not stop there. For a semantic task, measure mask precision/recall and calibration (ECE or reliability diagrams) against held-out point labels. For rendering, render identical camera paths before and after editing and report masked-region LPIPS/PSNR/SSIM only where a reference exists; separately report background leakage and disocclusion artifacts. Human review should score boundary quality, identity preservation, and failure severity. Every result should retain the manifest, source hash, renderer version, and camera set.

## Limitations and safety

PLY fields are not a universal Gaussian-splat contract. ASCII export can lose binary packing and comments from tools that rely on unusual extensions. The baseline does not understand mesh topology, depth ordering, view-dependent spherical harmonics, or semantic labels. A crop that looks correct from one orbit can still damage novel views. Thresholds should therefore be treated as an editing hypothesis, not ground truth. Do not use this baseline for safety-critical spatial measurement or to claim a model has learned object understanding.