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

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Wang, Xu, Wang, Zhiru, Xie, Shiyun, Pan, Chengwei, Chen, Yisong
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910160470933504
author Wang, Xu
Wang, Zhiru
Xie, Shiyun
Pan, Chengwei
Chen, Yisong
author_facet Wang, Xu
Wang, Zhiru
Xie, Shiyun
Pan, Chengwei
Chen, Yisong
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2604_21631
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle DualSplat: Robust 3D Gaussian Splatting via Pseudo-Mask Bootstrapping from Reconstruction Failures
Wang, Xu
Wang, Zhiru
Xie, Shiyun
Pan, Chengwei
Chen, Yisong
Computer Vision and Pattern Recognition
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.
title DualSplat: Robust 3D Gaussian Splatting via Pseudo-Mask Bootstrapping from Reconstruction Failures
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2604.21631