Decoupled Reprojection Consistency for Diagnosing 3D Gaussian Splatting Failures
Loading...
Date
2026
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
The Eurographics Association
Abstract
This paper introduces Decoupled Reprojection Consistency (DRC), a training-free diagnostic for 3D Gaussian Splatting (3DGS) that decomposes an ambiguous reprojection error into three interpretable maps: a geometry/visibility mismatch Edepth, a base photometric inconsistency Ebase computed under SH truncation (L=0), and a view-dependent residual Evd. By comparing cross-view reprojection errors from zeroth-order (base) and trained-degree (full) SH renders of the same model, DRC separates geometry/visibility-driven inconsistency from appearance-driven instability. Paired with a bivariate fingerprint (quadrant occupancy) and coverage reporting, DRC turns reprojection consistency into actionable per-pixel triage. The method is demonstrated on synthetic and real scenes, separating glossy regions from geometry failures in our examples.
Description
CCS Concepts: Computing methodologies → Rendering; Image-based rendering
@inproceedings{10.2312:egp.20261014,
booktitle = {Eurographics 2026 - Posters},
editor = {Gerrits, Tim and Teschner, Matthias},
title = {{Decoupled Reprojection Consistency for Diagnosing 3D Gaussian Splatting Failures}},
author = {Park, Jin-Hyeong },
year = {2026},
publisher = {The Eurographics Association},
ISSN = {1017-4656},
ISBN = {978-3-03868-300-1},
DOI = {10.2312/egp.20261014}
}