3D Gaussian Splatting & Novel View Synthesis
Robust Wide-Baseline 3D Gaussian Splatting Under Unreliable Foundation-Model Geometry Priors
Generated automatically from the limitations stated in 3 papers (ICLR, NeurIPS, ICCV), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
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.
Why it matters
Enables high-fidelity 3D Gaussian scene reconstruction from unposed, sparse, wide-baseline image sets without requiring slow classical Structure-from-Motion or continuous video trajectories.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Joint photometric and geometry uncertainty re-weighting: Implement an uncertainty-guided Gaussian densification scheme that weights MASt3R pointmaps and pose priors based on visual foundation model confidence and epipolar residuals; measure PSNR, SSIM, and geometric floater artifacts across baseline angles ranging from 15° to 90° on CO3D.
- 2
Differentiable rendering-loss pose refinement: Build a joint optimization loop where 3D Gaussian positions and upstream coarse camera poses are continuously refined against multi-view photometric consistency; measure camera rotation/translation error and novel view synthesis quality against COLMAP baselines.
- 3
Cycle-consistent graph rectification for foundation poses: Construct a relative pose graph from all pairwise foundation model outputs and solve for synchronized global camera trajectories prior to Gaussian initialization; measure pose accuracy and reconstruction completeness on sparse-view benchmarks.
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
Rapid architectural scaling in next-generation multi-view foundation models could resolve wide-baseline pose regression and depth estimation natively, dissolving the need for downstream pose-and-geometry correction within the Gaussian splatting pipeline.
Evidence
Each paper's own statement of the limitation, verbatim.
- Aligned Novel View Image and Geometry Synthesis via Cross-modal Attention InstillationICLR 2026
Relies on an external off-the-shelf geometry/pose predictor (e.g., VGGT) to obtain initial coordinate pointmaps and relative reference poses.
- Omnidirectional 3D Scene Reconstruction from Single ImageNeurIPS 2025
Relies on MASt3R pose-estimation quality, which the authors note degrades for large angular differences between paired views, constraining the sliding-window size and orbit stepping
- Bolt3D: Generating 3D Scenes in SecondsICCV 2025
Requires camera poses for inputs and target views, and depends on MASt3R-derived pseudo-ground-truth geometry, which carries residual noise and imperfections
Nearest existing work
- Deep Gaussian from Motion: Exploring 3D Geometric Foundation Models for Gaussian SplattingNeurIPS 2025
- FSFSplatter: Geometrically Accurate Reconstruction with Free Sparse-view Images within 2 minutesCVPR 2026
- BA-GS: Bayesian Adaptive Gaussian Splatting for SFM-Free 3D ReconstructionCVPR 2026
- No Pose at All: Self-Supervised Pose-Free 3D Gaussian Splatting from Sparse ViewsICCV 2025
- FreeSplatter: Pose-free Gaussian Splatting for Sparse-view 3D ReconstructionICCV 2025
- MetroGS: Efficient and Stable Reconstruction of Geometrically Accurate High-Fidelity Large-Scale ScenesCVPR 2026
- Liberated-GS: 3D Gaussian Splatting Independent from SfM Point CloudsICCV 2025
- PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting for Novel View SynthesisICML 2025
- COLMAP-Free 3D Gaussian SplattingCVPR 2024
- Sparfels: Fast Reconstruction from Sparse Unposed ImageryICCV 2025
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage TrainingNeurIPS 2024
- GaussFusion: Improving 3D Reconstruction in the Wild with A Geometry-Informed Video GeneratorCVPR 2026
- NopeRoomGS: Indoor 3D Gaussian Splatting Optimization without Camera Pose InputNeurIPS 2025
- Sparse2DGS: Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse ViewsCVPR 2025
- SparseSplat: Towards Applicable Feed-Forward 3D Gaussian Splatting with Pixel-Unaligned PredictionCVPR 2026
Related open problems
3D Gaussian Splatting & Novel View Synthesis
Cross-Domain Robustness and Generalization Benchmarking for 3D Gaussian Splatting
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.
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.