3D Gaussian Splatting & Novel View Synthesis
Cross-Domain Robustness and Generalization Benchmarking for 3D Gaussian Splatting
Generated automatically from the limitations stated in 7 papers (CVPR, ICLR, NeurIPS), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Novel view synthesis and 3D Gaussian Splatting methods are currently evaluated almost exclusively on room-scale indoor datasets such as Replica, ScanNet++, and Gibson. Because no single pipeline has been systematically evaluated across indoor, outdoor unbounded, panoramic, and object-centric domains, the robustness and failure modes of these representations under domain shift remain entirely unknown. Practitioners and researchers cannot determine whether observed performance gains are artifacts of bounded indoor geometries or if current densification and optimization heuristics translate to unconstrained environments.
Why it matters
Establishes the first standardized empirical baseline for domain generalization in 3D Gaussian Splatting, identifying concrete failure boundaries across scene types.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Unified cross-domain benchmark: Aggregate existing standardized indoor (ScanNet++, Replica) and outdoor/unbounded (e.g., Mip-NeRF 360, Tanks and Temples) datasets into a single evaluation harness, measuring rendering quality (PSNR, SSIM, LPIPS) and training stability across several state-of-the-art 3DGS variants under identical parameter regimes.
- 2
Zero-shot domain transfer evaluation: Evaluate feed-forward and feature-conditioned 3D Gaussian models trained solely on indoor scans directly on outdoor and panoramic scenes, measuring the rate of visual artifact degradation and depth consistency loss.
- 3
Initialization and heuristic stress testing: Systematically test how standard indoor Gaussian densification, pruning, and learning-rate schedules behave across varying scene scales and camera trajectories (e.g., inward-facing vs. unbounded forward-facing).
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
If simple parameter retuning (such as adjusting background color or bounding thresholds) trivially resolves cross-domain degradation, the work would reduce to routine hyperparameter tuning rather than revealing fundamental generalization limits.
Evidence
Each paper's own statement of the limitation, verbatim.
- econSG: Efficient and Multi-view Consistent Open-Vocabulary 3D Semantic GaussiansICLR 2025
Evaluated only on room-scale indoor benchmarks (Replica, ScanNet, LERF, 3DOVS); generalization to large outdoor scenes is untested
- Splatter-360: Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic ImagesCVPR 2025
Generalization is evaluated only on indoor scenes (HM3D, Replica); outdoor or large-scale panoramic scenes are untested
- Retri3D: 3D Neural Graphics Representation RetrievalICLR 2025
Evaluation is limited to indoor/room-scale datasets (LERF, ScanNet++); no results on outdoor, unbounded, or synthetic isolated-object scenes
- IndoorGS: Geometric Cues Guided Gaussian Splatting for Indoor Scene ReconstructionCVPR 2025
Only evaluated on 14 indoor scenes; no evidence on outdoor, object-centric, or large-scale unbounded environments
- Efficient Decoupled Feature 3D Gaussian Splatting via Hierarchical CompressionCVPR 2025
Evaluated only on 5 indoor Replica scenes with LSeg features; outdoor or open-vocabulary generalization is untested
- HiCoM: Hierarchical Coherent Motion for Dynamic Streamable Scenes with 3D Gaussian SplattingNeurIPS 2024
Evaluated only on indoor scenes; generalization to outdoor or more complex environments is unverified
- 3DFIRES: Few Image 3D REconstruction for Scenes with Hidden SurfacesCVPR 2024
Trained and evaluated only on indoor Gibson/Omnidata scenes with wide-baseline sampling; performance on other domains is untested
Nearest existing work
- MVSGaussian: Fast Generalizable Gaussian Splatting Reconstruction from Multi-View StereoECCV 2024
- HO-Gaussian: Hybrid Optimization of 3D Gaussian Splatting for Urban ScenesECCV 2024
- SplatSuRe: Selective Super-Resolution for Multi-view Consistent 3D Gaussian SplattingCVPR 2026
- Metropolis-Hastings Sampling for 3D Gaussian ReconstructionNeurIPS 2025
- SPC-GS: Gaussian Splatting with Semantic-Prompt Consistency for Indoor Open-World Free-view Synthesis from Sparse InputsCVPR 2025
- MetroGS: Efficient and Stable Reconstruction of Geometrically Accurate High-Fidelity Large-Scale ScenesCVPR 2026
- FreeSplat: Generalizable 3D Gaussian Splatting Towards Free View Synthesis of Indoor ScenesNeurIPS 2024
- GeoGaussian: Geometry-aware Gaussian Splatting for Scene RenderingECCV 2024
- QuickSplat: Fast 3D Surface Reconstruction via Learned Gaussian InitializationICCV 2025
- VastGaussian: Vast 3D Gaussians for Large Scene ReconstructionCVPR 2024
- PCGS: Deblurring 3D Gaussian Splatting with Patch ComparisonICML 2026
- GaussianPro: 3D Gaussian Splatting with Progressive PropagationICML 2024
- 3D Gaussian Splatting as Markov Chain Monte CarloNeurIPS 2024
- NopeRoomGS: Indoor 3D Gaussian Splatting Optimization without Camera Pose InputNeurIPS 2025
- GS^2: Graph-based Spatial Distribution Optimization for Compact 3D Gaussian SplattingCVPR 2026
Related open problems
3D Gaussian Splatting & Novel View Synthesis
Dynamic 3D Gaussian Splatting Under Non-Smooth and Discontinuous Motion Dynamics
Current dynamic novel view synthesis and 4D Gaussian Splatting frameworks assume smooth, continuous spatio-temporal deformations, linear velocities, and fixed temporal windows. When scenes exhibit abrupt impacts, rapid accelerations, chaotic object trajectories, or zooming and variable frame rates, these continuous deformation priors fail and yield severe geometric tearing, blurring, or total tracking collapse. This fundamentally blocks dynamic radiance fields from capturing real-world interactions such as sports action, mechanical impacts, and sudden dynamic occlusions.
3D Gaussian Splatting & Novel View Synthesis
Robust Wide-Baseline 3D Gaussian Splatting Under Unreliable Foundation-Model Geometry Priors
Modern feed-forward 3D Gaussian Splatting pipelines rely on pairwise foundation models (such as MASt3R or VGGT) to supply initial coordinate pointmaps and relative camera poses. Because these foundation predictors degrade significantly as angular view differences increase and inject persistent geometric noise, downstream view synthesis fails when views are not captured in dense, small-step sequences. Treating these imperfect upstream predictions as rigid pseudo-ground truth structurally blocks feed-forward novel view synthesis on sparse, wide-baseline image collections.