SSA-3DGS: Unsupervised Removal of Screen-Space Artifacts for 3D Gaussian Splatting

Jul 16, 2026

Upload by Noah

Novel View Synthesis pipelines such as 3D Gaussian Splatting depend on clean, multi-view consistent posed images. Real-world captures often violate this assumption due to screen-space artifacts that are fixed to the 2D image plane—physical defects (e.g., dead pixels, dust), environmental obstructions (rain or mud on a lens enclosure), capture occlusions (fingers, dashboard views), and digital overlays (watermarks, UI elements). These artifacts are incorrectly baked into 3D as near-camera floaters, degrading novel-view rendering. SSA-3DGS introduces an unsupervised dual-objective that jointly recovers an artifact-free 3D scene and a learnable 2D overlay shared across views. Exploiting geometric consensus and motion parallax, the method segments artifacts without ground-truth masks, improving reconstruction fidelity by up to 9 dB PSNR over 3DGS trained on corrupted inputs while preserving the corrupting artifacts’ appearance.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

Introduction

Problem: Screen-space artifacts in posed multi-view inputs

Limitations of existing restoration and watermark-removal approaches

Proposed solution: SSA-3DGS framework

Evaluation and results

SSA-3DGS: Unsupervised Removal of Screen-Space Artifacts for 3D Gaussian Splatting Kristof Overdulve Lode Jorissen Nick Michiels Digital Future Lab, Flanders Make, Hasselt University {kristof.overdulve, lode.jorissen, [nick.michiels](nick.michiels}@uhasselt.be)[}](nick.michiels}@uhasselt.be)[@uhasselt.be](nick.michiels}@uhasselt.be) arXiv :2607 .05598v 1 [ cs .GR] 6 Jul 2026 Our Method Clean renders (PSNR val. set: 29.48 dB) Reconstructed artifact Input: Corrupted posed images Baseline (3DGS) Renders containing artifacts (PSNR val. set: 21.51 dB) . Figure 1 . Unsupervised decomposition of screen-space artifacts. From a set of posed images corrupted by static overlays (e.g., watermarks, dirt on the lens enclosure, . . . ), SSA-3DGS recovers a high-fidelity 3D reconstruction and segments the artifact (top right) without ground-truth masks or manual supervision. Abstract Novel View Synthesis (NVS) methods, such as 3D Gaussian Splatting (3DGS), rely heavily on the assumption of clean, multi-view consistent, posed input images. Realworld captures can violate this assumption due to screenspace artifacts—static occlusions fixed to the 2D image plane rather than to the 3D world. Common examples include physical sensor defects, environmental obstructions (such as rain or mud on the lens enclosure), capture obstructions (such as a thumb over the camera sensor or a dashboard visible in dashcam footage), and digital overlays (such as watermarks or UI elements) . When present, they are erroneously baked into the 3D geometry as “floaters” or near-camera artifacts, degrading the quality of novel-view rendering. In this work, we propose SSA-3DGS, an unsupervised framework that jointly optimizes a 3D scene anda learnable 2D overlay to recover a clean 3D scene and the corrupting artifacts. By exploiting geometric consensus across views, our method effectively disentangles static artifacts from the 3D scene geometry without supervision or manual input. Across diverse synthetic corruptions anda self-captured real-world dataset, SSA-3DGS improves reconstruction fidelity by up to 9 dB PSNR over 3DGS trained on the same corrupted inputs, while faithfully preserving the corrupting artifact. 1. Introduction The goal of Novel View Synthesis (NVS) is to capture the visual essence of a real-world scene from multiple viewpoints and render it from arbitrary novel angles. This field has witnessed a paradigm shift with the advent of Neural Radiance Fields (NeRF) [13] and, more recently, 3D Gaussian Splatting (3DGS) [9] . 3DGS enables high-fidelity, realtime rendering by representing scenes as a collection of explicit 3D Gaussians rather than implicit neural weights, making it a prime candidate for real-time applications and 3D digital content creation. However, the reconstruction quality of novel-view synthesis algorithms relies heavily on a fundamental assump- tion in Multi-View Stereo (MVS): that the input images provide a multi-view-consistent representation of the underlying scene geometry. While recent research has effectively modeled world-space inconsistencies such as illumination changes [15] or transient moving elements like pedestrians [12, 16, 26, 27], these approaches struggle to handle screenspace artifacts (see Fig. 2 for example screen-space artifacts) . Distinct from world-space objects, these artifacts are static relative to the camera sensor, manifesting as physical defects (dead pixels or dust), environmental effects (rain or mud spatters on the lens enclosure), capture obstructions (visible fingers, or car dashboards in dashcam footage), or digital overlays (watermarks, timestamps, or UI elements) . Because these elements are fixed to the image plane, they effectively “move” relative to the world geometry as the camera moves. Consequently, standard novel view synthesis and even approaches focusing on world-space inconsistencies often fail to reject them, erroneously baking the artifacts into the 3D scen

FAQs

What are screen-space artifacts, and why do they harm 3D Gaussian Splatting?

Screen-space artifacts are elements static relative to the camera sensor and fixed to the 2D image plane. Because they move differently than world geometry, standard NVS methods mistakenly incorporate them into the 3D scene, producing near-camera floaters and degraded novel-view rendering.

How does SSA-3DGS remove artifacts without supervision?

SSA-3DGS jointly optimizes 3D scene parameters and a learnable 2D overlay that models the corrupting artifacts. It uses motion parallax across views to decouple screen-space artifacts from scene geometry, enabling artifact segmentation and clean reconstruction without ground-truth masks or manual input.

What improvements does SSA-3DGS achieve compared with baseline 3DGS on corrupted inputs?

On diverse synthetic corruptions and a self-captured real-world dataset, SSA-3DGS increases reconstruction fidelity by up to 9 dB PSNR over 3DGS trained on the same corrupted inputs, while preserving the corrupting artifact appearance faithfully.

Jul 16, 2026

Upload by Noah