Home | Publications | GKG+26

Perceptual Quality Densification for Gaussian Splatting

MCML Authors

Link to Profile Björn Ommer

Björn Ommer

Prof. Dr.

Core PI

Computer Vision & Learning

Abstract

Abstract. We propose a perceptually guided densification framework for dense multi-view 3D reconstruction with 3D Gaussian Splatting (3DGS). Accurate and visually faithful 3D reconstruction is central to applications such as robotics, AR/VR, digital preservation, and immersive media. Yet, standard 3DGS pipelines rely on heuristic densification criteria driven mainly by pixel-space losses such as L1 and SSIM, which correlate only imperfectly with human perception. As a result, Gaussians may be allocated ineffciently, and perceptually important regions such as low-contrast areas, repetitive textures, and fine structures can remain under-reconstructed. Our method introduces three components: a learnable 2D perceptual quality estimator that predicts dense artifact-aware confidence maps, a multi-view aggregation strategy that converts per-view estimates into Gaussian-level scores, and a perceptually guided densification rule that selectively refines unreliable regions. Experiments show that our approach improves visual fidelity and reduces artifacts while preserving strong reconstruction accuracy. This confirms that perceptual feedback is an effective signal for guiding model capacity in 3DGS.

inproceedings GKG+26

GCPR 2026

German Conference on Pattern Recognition. Siegen, Germany, Sep 22-25, 2026. To be published.

Authors

O. Kotovenko • M. Gui • J. Schusterbauer • B. Ommer

Research Area

 B1 | Computer Vision

BibTeXKey: GKG+26

#p-ommer
Back to Top