gaussian splatting: A Multiverse of 3D Innovation, from Surgical Reconstruction to Digital Twins
Latest 40 papers on gaussian splatting: Feb. 28, 2026
Step into the exciting realm of 3D Gaussian Splatting (3DGS), a technology rapidly reshaping how we perceive, reconstruct, and interact with digital worlds. Once a niche technique, 3DGS has exploded into an area of intense research, promising real-time, high-fidelity 3D scene representation with unprecedented efficiency. Recent breakthroughs, as showcased in a collection of cutting-edge papers, are pushing the boundaries further, tackling everything from dynamic scenes and medical imaging to robust scene understanding and digital twin generation.
The Big Idea(s) & Core Innovations:
At its heart, 3DGS represents scenes as a collection of 3D Gaussians, each with properties like position, scale, rotation, and color. This simple yet powerful representation allows for incredibly fast and high-quality rendering. The papers we’re exploring illustrate a fascinating convergence of ideas: enhancing traditional 3DGS, extending it to 4D (space-time) dynamics, and applying it to complex real-world challenges.
Several works are focused on improving the core 3DGS process. For instance, GIFSplat: Generative Prior-Guided Iterative Feed-Forward 3D Gaussian Splatting from Sparse Views by researchers from La Trobe University and Cisco Research, introduces an iterative feed-forward framework that leverages generative priors to achieve high-quality reconstructions from sparse views, significantly improving PSNR without test-time optimization. Similarly, RAP: Fast Feedforward Rendering-Free Attribute-Guided Primitive Importance Score Prediction for Efficient 3D Gaussian Splatting Processing from Shanghai Jiao Tong University and the University of Missouri–Kansas City, proposes a rendering-free method to predict Gaussian importance scores directly from intrinsic attributes, leading to more efficient pruning and compression.
Robustness and generalization are also major themes. DefenseSplat: Enhancing the Robustness of 3D Gaussian Splatting via Frequency-Aware Filtering by Case Western Reserve University, tackles the critical problem of adversarial attacks on 3DGS by employing frequency-aware filtering, improving model robustness without clean ground-truth data. In a similar vein, Distractor-free Generalizable 3D Gaussian Splatting from Nanjing University and City University of Hong Kong, proposes DGGS to eliminate distractors during training and inference, leading to more stable and artifact-free reconstructions that generalize across scenes.
Extending 3DGS to handle dynamic and complex environments is another key advancement. Latent Gaussian Splatting for 4D Panoptic Occupancy Tracking from the University of Freiburg introduces LaGS, a unified framework that combines geometric reconstruction and semantic understanding for state-of-the-art 4D panoptic occupancy tracking. For challenging aerial scenarios, AeroDGS: Physically Consistent Dynamic Gaussian Splatting for Single-Sequence Aerial 4D Reconstruction by The Ohio State University, uses physics-guided optimization to achieve stable 4D reconstruction from monocular aerial videos. In a more theoretical exploration, DARB-Splatting: Generalizing Splatting with Decaying Anisotropic Radial Basis Functions from the University of Moratuwa and the University of Adelaide, generalizes splatting functions beyond Gaussians, showing that other radial basis functions can offer faster convergence and lower memory usage.
Medical applications are seeing significant advancements as well. RU4D-SLAM: Reweighting Uncertainty in Gaussian Splatting SLAM for 4D Scene Reconstruction from Capital Normal University and Saarland University, introduces a robust 4D Gaussian splatting SLAM that handles motion blur and integrates uncertainty-aware perception, particularly useful in dynamic medical imaging. This is echoed by 4D Monocular Surgical Reconstruction under Arbitrary Camera Motions and NRGS-SLAM: Monocular Non-Rigid SLAM for Endoscopy via Deformation-Aware 3D Gaussian Splatting, both demonstrating high-quality 4D reconstruction of deformable surgical scenes from monocular endoscopic videos, critical for improving surgical navigation.
Beyond reconstruction, 3DGS is enabling novel applications. BrepGaussian: CAD reconstruction from Multi-View Images with Gaussian Splatting by Nanjing University and Nanjing Bridge Intelligent Management Co.,Ltd., showcases the first framework to reconstruct complete B-rep CAD models directly from multi-view images without point cloud supervision. Meanwhile, WildGHand: Learning Anti-Perturbation Gaussian Hand Avatars from Monocular In-the-Wild Videos by MIT, Stanford University, and Google Research, generates realistic and robust hand avatars from challenging in-the-wild videos, opening doors for advanced human-computer interaction.
Under the Hood: Models, Datasets, & Benchmarks:
These innovations are powered by novel architectures and rigorously evaluated on challenging datasets:
- LaGS (Latent Gaussian Splatting for 4D Panoptic Occupancy Tracking): Leverages latent Gaussian splatting on Occ3D nuScenes and Waymo datasets for 4D panoptic occupancy tracking. Code and resources available at https://lags.cs.uni-freiburg.de/.
- PackUV & PackUV-2B (PackUV: Packed Gaussian UV Maps for 4D Volumetric Video): Introduces a compact 4D Gaussian representation encoding volumetric data into 2D UV maps and a massive 4D multi-view dataset with over 2 billion frames, setting a new benchmark for volumetric video research. Resources at https://ivl.cs.brown.edu/packuv.
- GSTurb (GSTurb: Gaussian Splatting for Atmospheric Turbulence Mitigation): Combines Gaussian splatting with optical-flow fields for turbulence mitigation in both synthetic and real-world conditions. Code available at https://github.com/Duhl/Liamz/3DGS_turbulence/tree/main.
- BetterScene (BetterScene: 3D Scene Synthesis with Representation-Aligned Generative Model): Enhances novel view synthesis using Stable Video Diffusion and vision foundation model-aligned representations with 3D Gaussian Splatting as an input renderer. Code at https://github.com/USACE-ERDC-GRL/BetterScene.
- SwiftNDC (SwiftNDC: Fast Neural Depth Correction for High-Fidelity 3D Reconstruction): A neural depth correction framework for high-fidelity 3D reconstruction, improving depth maps for mesh reconstruction and novel view synthesis. Resources at https://arxiv.org/pdf/2602.22565.
- AeroDGS & Aero4D (AeroDGS: Physically Consistent Dynamic Gaussian Splatting for Single-Sequence Aerial 4D Reconstruction): A 4D Gaussian Splatting framework with physics-guided optimization for aerial reconstruction, introduces the real-world Aero4D dataset. Resources at https://arxiv.org/pdf/2602.22376.
- DGGS (Distractor-free Generalizable 3D Gaussian Splatting): A distractor-free generalizable 3DGS training and inference paradigm leveraging 3D consistency and semantic priors. Resources at https://arxiv.org/pdf/2411.17605.
- MoGaF (Space-Time Forecasting of Dynamic Scenes with Motion-aware Gaussian Grouping): A framework for long-term scene extrapolation using motion-aware Gaussian grouping and group-wise optimization. Code available at https://slime0519.github.io/mogaf.
- DAGS-SLAM (DAGS-SLAM: Dynamic-Aware 3DGS SLAM via Spatiotemporal Motion Probability and Uncertainty-Aware Scheduling): Enhances dynamic-aware 3DGS SLAM with spatiotemporal motion probability and uncertainty-aware scheduling. Code at https://github.com/your-username/DAGS-SLAM.
- BrepGaussian (BrepGaussian: CAD reconstruction from Multi-View Images with Gaussian Splatting): Reconstructs B-rep CAD models from multi-view images using 2D Gaussian Splatting and primitive fitting. Resources at https://arxiv.org/pdf/2602.21105.
- DropAnSH-GS (Dropping Anchor and Spherical Harmonics for Sparse-view Gaussian Splatting): A Dropout strategy for 3DGS to combat overfitting in sparse-view conditions. Resources at https://sk-fun.fun/DropAnSH-GS.
- RU4D-SLAM (RU4D-SLAM: Reweighting Uncertainty in Gaussian Splatting SLAM for 4D Scene Reconstruction): Integrates uncertainty-aware perception with 4D Gaussian splatting for dynamic scene reconstruction. Code and resources at https://ru4d-slam.github.io.
- WildGHand (WildGHand: Learning Anti-Perturbation Gaussian Hand Avatars from Monocular In-the-Wild Videos): Generates robust hand avatars from monocular in-the-wild videos. Code available at https://github.com/XuanHuang0/WildGHand.
- PUN (Peering into the Unknown: Active View Selection with Neural Uncertainty Maps for 3D Reconstruction): An active view selection method using neural uncertainty maps (UPNet) for efficient 3D reconstruction. Code at https://github.com/ZhangLab-DeepNeuroCogLab/PUN.
- Augmented Radiance Field (Augmented Radiance Field: A General Framework for Enhanced Gaussian Splatting): Enhances Gaussian Splatting by modeling specular highlights through view-dependent opacity. Code at https://github.com/xiaoxinyyx/augs.
- One2Scene (One2Scene: Geometric Consistent Explorable 3D Scene Generation from a Single Image): Generates explorable 3D scenes from a single image using feed-forward 3D Gaussian Splatting. Code available at https://one2scene5406.github.io/.
- DefenseSplat (DefenseSplat: Enhancing the Robustness of 3D Gaussian Splatting via Frequency-Aware Filtering): A frequency-aware defense strategy for 3DGS against adversarial attacks. Resources at https://arxiv.org/pdf/2602.19323.
- Diff2DGS (Diff2DGS: Reliable Reconstruction of Occluded Surgical Scenes via 2D Gaussian Splatting): Uses 2D Gaussian splatting for reliable reconstruction of occluded surgical scenes. Code available at https://anonymous.4open.science/r/Diff2DGS.
- Local-EndoGS (4D Monocular Surgical Reconstruction under Arbitrary Camera Motions): A framework for high-quality 4D surgical reconstruction from monocular endoscopic videos. Code at https://github.com/IRMVLab/Local-EndoGS.
- NRGS-SLAM (NRGS-SLAM: Monocular Non-Rigid SLAM for Endoscopy via Deformation-Aware 3D Gaussian Splatting): A monocular non-rigid SLAM approach using deformation-aware 3D Gaussian splatting for endoscopy. Resources at https://arxiv.org/pdf/2602.17182.
- B3-Seg (B3-Seg: Camera-Free, Training-Free 3DGS Segmentation via Analytic EIG and Beta-Bernoulli Bayesian Updates): A camera-free, training-free method for open-vocabulary 3DGS segmentation. Resources at https://arxiv.org/pdf/2602.17134.
- Multimodal Gaussian Splatting (3D Scene Rendering with Multimodal Gaussian Splatting): Integrates multiple data sources for enhanced 3D scene rendering. Code at https://github.com/graphdeco-inria/gaussian-splatting.
- i-PhysGaussian (i-PhysGaussian: Implicit Physical Simulation for 3D Gaussian Splatting): Combines 3DGS with an implicit Material Point Method integrator for stable physical simulations. Code available at https://github.com/sydneyai/i-physgaussian.
- GS-based Digital Twins (Three-dimensional Damage Visualization of Civil Structures via Gaussian Splatting-enabled Digital Twins): A method for 3D damage visualization in civil structures using Gaussian Splatting. Code available at https://github.com/shuow2/Gaussian-Splatting-enabled-digital-twins-for-3D-damage-visualization-of-civil-structures.
- MeGA (MeGA: Hybrid Mesh-Gaussian Head Avatar for High-Fidelity Rendering and Head Editing): A hybrid mesh-Gaussian representation for human head avatars, using the NeRSemble dataset. Resources at https://arxiv.org/pdf/2404.19026.
- 3DGEER (3DGEER: 3D Gaussian Rendering Made Exact and Efficient for Generic Cameras): Achieves exact geometric rendering of 3D Gaussians under generic camera models with Particle Bounding Frustum (PBF) and Bipolar Equiangular Projection (BEAP). Resources at https://zixunh.github.io/3d-geer.
- Semantic-Guided 3D Gaussian Splatting (Semantic-Guided 3D Gaussian Splatting for Transient Object Removal): Uses CLIP-based semantic filtering for transient object removal, evaluated on the RobustNeRF benchmark. Resources at https://arxiv.org/pdf/2602.15516.
- DAV-GSWT (DAV-GSWT: Diffusion-Active-View Sampling for Data-Efficient Gaussian Splatting Wang Tiles): Combines diffusion priors and active view sampling for data-efficient Gaussian Splatting Wang Tiles. Code available at https://github.com/DAV-GSWT/DAV-GSWT.
- Time-Archival Camera Virtualization (Time-Archival Camera Virtualization for Sports and Visual Performances): A framework for camera virtualization and dynamic scene rendering using neural implicit representations. Resources at https://arxiv.org/pdf/2602.15181.
Impact & The Road Ahead:
The implications of these advancements are profound. From revolutionizing autonomous driving with robust 4D perception (LaGS) to enabling highly accurate surgical navigation with real-time deformable tissue reconstruction (Local-EndoGS, NRGS-SLAM, Diff2DGS), 3DGS is proving to be a versatile powerhouse. Its ability to create explorable 3D scenes from a single image (One2Scene) and generate photorealistic large-scale outdoor scenes from UAV imagery (Large-scale Photorealistic Outdoor 3D Scene Reconstruction from UAV Imagery Using Gaussian Splatting Techniques) will transform industries like urban planning, virtual tourism, and film production.
The advent of digital twins for civil structures capable of dynamic damage visualization (Three-dimensional Damage Visualization of Civil Structures via Gaussian Splatting-enabled Digital Twins) signals a new era for infrastructure monitoring. Furthermore, tools like B3-Seg (B3-Seg: Camera-Free, Training-Free 3DGS Segmentation via Analytic EIG and Beta-Bernoulli Bayesian Updates) promise fast, interactive 3D asset editing, streamlining creative workflows in game development and visual effects.
The future of 3DGS lies in pushing the boundaries of realism, efficiency, and real-world applicability. Expect to see further integration of semantic understanding, even more robust handling of dynamic environments, and the exploration of novel mathematical functions to represent and render scenes. The drive for efficient processing, as seen in RAP and PUN, will remain critical for real-time applications. Gaussian splatting isn’t just a rendering technique; it’s a foundational shift in how we build and interact with digital realities, and these papers are charting an exciting course forward.
Share this content:
Post Comment