Real-time high-fidelity dense reconstruction in unconstrained environments

Establish visual SLAM methods that achieve high-fidelity, real-time dense 3D reconstruction in unconstrained environments.

Background

The paper surveys traditional visual SLAM approaches, distinguishing between sparse and dense methods, and highlights progress such as KinectFusion, BundleFusion, and MASt3R-SLAM that improve robustness and scalability. Despite these advances, the authors explicitly note that attaining both real-time performance and high-fidelity dense reconstruction across unconstrained, real-world scenarios remains unresolved.

VBGS-SLAM is introduced to address fragility and lack of uncertainty modeling in current Gaussian Splatting-based SLAM by formulating a probabilistic, variational inference framework. While the proposed method demonstrates strong results on several datasets, the broader goal of universally achieving high-fidelity, real-time dense reconstruction in unconstrained environments is identified as an open challenge.

References

Despite these advances, achieving high-fidelity, real-time dense reconstruction in unconstrained environments remains an open challenge.

— VBGS-SLAM: Variational Bayesian Gaussian Splatting Simultaneous Localization and Mapping  (2604.02696 - Zhu et al., 3 Apr 2026) in Section 2.1 Traditional Visual SLAM (Related Works)

Despite these significant advancements, scaling streaming 3D perception to unconstrained, ultra-large outdoor environments with dynamic occlusions remains an open challenge. Future research will explore adapting our Stream3Dv2 framework to outdoor scene understanding and further coupling it with physical embodied agents for real-time robotic manipulation.

— Stream3Dv2: Geometric-Semantic Fusion Enhanced Streaming Zero-Shot 3D Scene Understanding  (2608.21136 - Xu et al., 21 Aug 2026) in Section Conclusion