Robust referring segmentation under imperfect 3DGS reconstruction

Improve the robustness of open-vocabulary referring segmentation in 3D Gaussian Splatting when the reconstructed Gaussian representation is noisy, incomplete, or inaccurate because of occlusion, transparent objects, reflective surfaces, or insufficient views.

Background

QAGaussian operates directly on reconstructed 3D Gaussian Splatting scenes, so errors in the underlying geometric and appearance representation can propagate to Gaussian-level language-guided masks. The authors identify several reconstruction conditions—including occlusion, transparency, reflectance, and insufficient viewing coverage—that can degrade the resulting segmentation, leaving robustness to imperfect scene representations unresolved.

References

First, the method depends on the quality of the reconstructed 3DGS scene. If the Gaussian representation is noisy, incomplete, or inaccurate due to occlusion, transparent objects, reflective surfaces, or insufficient views, the predicted mask may also be degraded.

— Beyond Similarity Matching: Structured Reasoning for Open-Vocabulary Referring Segmentation in 3DGS  (2608.16103 - Wang et al., 17 Aug 2026) in Section Discussion, subsection “Limitations and future work”

Our experiments isolate semantic errors by constructing maps with ground-truth depth and camera poses; robustness to geometric errors, therefore, remains to be evaluated.

— VoxelFix: Post-Hoc Semantic Correction of Completed 3D Voxel Maps  (2609.05114 - Sundarasami et al., 4 Sep 2026) in Section Conclusion

To the best of our knowledge, direct 3D point-cloud segmentation of infrastructure at tunnel scene scale remains an open challenge for current foundation models.

— R4Tun: LLM-guided adaptive segmental tunnel lining segmentation in point clouds  (2609.11360 - Tao et al., 10 Sep 2026) in Section 2.2, “Foundation-model pipelines”