DualSplat: Robust 3D Gaussian Splatting via Pseudo-Mask Bootstrapping from Reconstruction Failures

TL;DR

DualSplat introduces a novel framework that leverages reconstruction failures to generate pseudo-masks, enabling robust 3D Gaussian Splatting in scenes with transient objects and improving performance over existing methods.

Contribution

The paper proposes DualSplat, a failure-to-prior approach that converts reconstruction failures into explicit priors for improved transient object handling in 3D Gaussian Splatting.

Findings

DualSplat outperforms existing baselines on RobustNeRF and NeRF On-the-go.

It shows significant improvements in scenes with transient objects.

Pseudo-masks guide a second-pass optimization for cleaner reconstructions.

Abstract

While 3D Gaussian Splatting (3DGS) achieves real-time photorealistic rendering, its performance degrades significantly when training images contain transient objects that violate multi-view consistency. Existing methods face a circular dependency: accurate transient detection requires a well-reconstructed static scene, while clean reconstruction itself depends on reliable transient masks. We address this challenge with DualSplat, a Failure-to-Prior framework that converts first-pass reconstruction failures into explicit priors for a second reconstruction stage. We observe that transients, which appear in only a subset of views, often manifest as incomplete fragments during conservative initial training. We exploit these failures to construct object-level pseudo-masks by combining photometric residuals, feature mismatches, and SAM2 instance boundaries. These pseudo-masks then guide a clean second-pass 3DGS optimization, while a lightweight MLP refines them online by gradually shifting from prior supervision to self-consistency. Experiments on RobustNeRF and NeRF On-the-go show that DualSplat outperforms existing baselines, demonstrating particularly clear advantages in transient-heavy scenes and transient regions.