滴滴和港中文最新的前馈3D重建算法UniSplat！史少帅参与~

Core Viewpoint - The article discusses the introduction of UniSplat, a novel feed-forward framework for dynamic scene reconstruction in autonomous driving, which effectively integrates spatio-temporal information from multi-camera video inputs to enhance the robustness and quality of 3D scene reconstruction [6][44]. Background Review - Reconstructing 3D scenes from urban driving scenarios has become a core capability for autonomous driving systems, supporting critical tasks such as simulation, scene understanding, and long-horizon planning [5]. - Recent advancements in 3D Gaussian splatting technology have shown impressive rendering efficiency and fidelity, but existing methods often assume significant overlap between input images and rely on scene-by-scene optimization, limiting their applicability in real-time driving scenarios [5][6]. UniSplat Overview - UniSplat is designed to address the challenges of robust reconstruction in dynamic driving scenes by constructing a unified 3D Scaffold that integrates multi-view spatial information and multi-frame temporal information [6][9]. - The framework operates in three stages: building a 3D Scaffold from multi-view images, performing spatio-temporal fusion, and decoding the fused Scaffold into Gaussian primitives [6][9]. Experimental Results - Evaluations on the Waymo Open dataset and NuScenes dataset demonstrate that UniSplat achieves state-of-the-art performance in both input view reconstruction and new view synthesis tasks, showcasing strong robustness and superior rendering quality even for views outside the original camera coverage [7][34]. - In the Waymo dataset, UniSplat outperforms existing methods such as MVSplat and DepthSplat across all metrics, achieving a PSNR of 25.37 and an SSIM of 0.765 [34]. - The model effectively distinguishes between dynamic and static elements in scenes, successfully mitigating ghosting artifacts during scene completion [40]. Methodology Details - The 3D Scaffold is constructed by inferring geometric structures using a geometric backbone model and supplementing it with semantic information from a visual backbone model [14][16]. - A dual-branch decoder is employed to generate dynamic-aware Gaussian primitives, enhancing detail retention and scene completeness [23][27]. - The framework incorporates a memory mechanism to accumulate static Gaussian representations over time, facilitating long-term scene completion [29][31]. Conclusion - UniSplat represents a significant advancement in the field of dynamic scene understanding and interactive 4D content creation, providing a robust foundation for future research in lifelong world modeling and autonomous driving applications [44].